Financial risk mathematics begins with uncertainty: several outcomes are possible, they do not have the same likelihood, and the consequences of being wrong are not symmetric. Probability, random variables, distributions, expectation, variance, conditional probability, Bayes’ rule, quantiles, tail risk, simulation and dependence give us a language for describing that uncertainty. This flagship guide builds the foundation beneath credit-risk models, market-risk models, portfolio mathematics, stress testing and quantitative finance without collapsing those specialist subjects into one page.
For readers searching for financial risk mathematics, probability in finance, probability distributions, expected value, variance, standard deviation, normal distribution, lognormal distribution, conditional probability, Bayes theorem, Monte Carlo simulation, Value at Risk, VaR, Expected Shortfall, tail risk, loss distributions, scenario analysis or financial risk modelling, the core lesson is that risk is a distribution, not one number. A mean, volatility, VaR or Expected Shortfall is a summary of that distribution under assumptions.
Current professional and regulatory material makes the same distinction. CFA Institute’s 2026 quantitative-methods curriculum develops expected values, variances, probability trees, conditional expectations, Bayes’ formula and simulation as foundations for investment analysis. The Basel Framework defines VaR as a loss threshold at a chosen confidence/horizon and Expected Shortfall as the average loss beyond the VaR threshold, with the market-risk framework using ES under specified regulatory conditions. This page explains the mathematics; it is not financial advice and it is not a substitute for a bank’s regulatory model.
50-Second Router
- Probability: a numerical model of uncertainty, governed by consistent rules.
- Random variable: a numerical quantity whose value depends on the outcome of an uncertain process.
- Expected value: probability-weighted average outcome; useful but not a guarantee.
- Variance and standard deviation: dispersion around the mean; powerful but symmetric.
- Conditional probability: probability after restricting attention to information or an event.
- Bayes’ rule: a coherent way to update probabilities when new evidence arrives.
- Distribution: the full map of possible values and their probabilities or densities.
- Quantile/VaR: a threshold; it says little by itself about severity beyond the threshold.
- Expected Shortfall: a tail-average measure that addresses severity beyond a VaR cutoff under the chosen model.
- Simulation: generate many scenarios from a probability model, transform them through a valuation system and study the resulting output distribution.
- Dependence: joint outcomes matter; correlation is useful but does not fully describe tail co-movement.
- Verification: probabilities must sum correctly, models should be calibrated and backtested, and assumptions should be stress-tested.
The Central Proposition: Risk Is a Distribution Before It Is a Metric
A single risk number is convenient because organisations need limits, reports and decisions. But the number is downstream of a distribution. If two portfolios have the same standard deviation but one has a much heavier left tail, they do not have the same loss profile. If two loans have the same expected loss but one defaults rarely with catastrophic severity while the other loses small amounts frequently, their operational and capital implications can differ.
The first job of financial risk mathematics is therefore to define the uncertain object. What can vary? What outcomes are possible? What information set is available? What horizon matters? Which probabilities are being modelled: historical frequencies, subjective beliefs, physical probabilities, risk-neutral probabilities or regulatory scenario weights?
Adrian’s discipline is to refuse an unlabeled probability. “There is a 5% chance” is incomplete until the event, horizon, conditioning information and model are named.
1. Sample space
Sample space is the set of possible outcomes in a probability model. It defines what the model is capable of representing. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. Write Ω for the outcome space and events as subsets of Ω. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. If a material outcome is absent from the sample space, no later probability formula can recover it. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into scenario design. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
2. Event
Event is a set of outcomes to which probability is assigned. Financial questions are often event questions: default, loss beyond a threshold, rate increase, covenant breach. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A) lies between 0 and 1. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Vague event definitions create untestable probabilities. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into risk limits. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
3. Probability axioms
Probability axioms is the consistency rules requiring nonnegative probabilities, total probability one and additivity for mutually exclusive events. They prevent incoherent uncertainty arithmetic. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(Ω)=1 and P(A∪B)=P(A)+P(B) for disjoint A,B. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Assigning overlapping event probabilities as though events were disjoint double counts risk. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into probability trees. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
4. Complement rule
Complement rule is the relation between an event and its non-occurrence. It often simplifies tail calculations. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A^c)=1−P(A). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Confusing ‘not loss’ with ‘profit’ can be wrong when zero or intermediate outcomes exist. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into default/nondefault modelling. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
5. Union probability
Union probability is the chance that at least one of several events occurs. It appears in operational incidents and multi-name default questions. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A∪B)=P(A)+P(B)−P(A∩B). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Adding probabilities without subtracting overlap overstates combined risk. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into joint events. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
6. Intersection probability
Intersection probability is the probability that events occur together. Joint stress often depends on intersections. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A∩B)=P(A|B)P(B). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Assuming intersection equals product silently assumes independence. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into dependence. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
7. Conditional probability
Conditional probability is probability of A given information/event B. It formalises how information changes the relevant population. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A|B)=P(A∩B)/P(B). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Using unconditional default rates for a stressed borrower state can be misleading. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into credit risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
8. Independence
Independence is a relationship where knowing one event does not change the probability of another. It is much stronger than zero correlation in general. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A∩B)=P(A)P(B). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Assuming independence for convenience can severely understate joint tail events. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into portfolio and credit risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
9. Bayes’ rule
Bayes’ rule is a relationship that updates a prior probability using evidence likelihoods. It is one of the cleanest mathematical models of learning from new information. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A|B)=P(B|A)P(A)/P(B). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Ignoring base rates can produce dramatic false-positive errors. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into fraud and credit signals. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
10. Law of total probability
Law of total probability is a method of combining conditional probabilities across mutually exclusive exhaustive states. It converts scenario-conditional risks into an unconditional probability. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(A)=ΣP(A|S_i)P(S_i). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Scenario weights must sum to one and states must be exhaustive for the stated partition. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into macro credit scenarios. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
11. Random variable
Random variable is a numerical function of an uncertain outcome. It turns events into quantities such as return, loss, default count or interest rate. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. X:Ω→R. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. A model of return is not automatically a model of price, wealth or loss without transformation. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into distribution modelling. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
12. Discrete random variable
Discrete random variable is a variable taking countable values. Defaults, ratings and transaction counts are common discrete objects. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. Use probability mass function p(x)=P(X=x). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly.
Failure mode. Applying continuous-density formulas to discrete events can misstate probabilities. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation.
Connection. The concept feeds into binomial and Poisson models. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage.
13. Continuous random variable
Continuous random variable is a variable represented with a density over intervals. Returns, rates and losses are often approximated continuously. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions.
Mathematics. P(a Failure mode. For a continuous model P(X=x)=0 at an exact point, despite positive density. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into normal and lognormal models. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Probability mass function is the probabilities attached to values of a discrete variable. It fully specifies a discrete distribution. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Σp(x)=1. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Negative masses or totals not equal to one invalidate the distribution. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into discrete finance. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Probability density function is a function whose integral over a region gives probability. It describes continuous distributions. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. f(x)≥0 and ∫f(x)dx=1. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Density can exceed one locally; it is not itself a point probability. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into continuous finance. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Cumulative distribution function is the probability that a variable is at or below a threshold. It unifies discrete and continuous distributions. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. F(x)=P(X≤x). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Mixing loss and return sign conventions can reverse tail interpretation. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into quantiles. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Quantile is a value below which a chosen fraction of the distribution lies. It turns a distribution into threshold risk measures. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. q_α=F^{-1}(α) under a chosen convention. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Discrete distributions may have nonunique interval-style quantile definitions. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into VaR. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Expected value is the probability-weighted average outcome. It is the centre of many valuation and risk calculations. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. E[X]=Σxp(x) or ∫xf(x)dx. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. An expectation can be an outcome that never actually occurs and can be dominated by rare tails. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into expected loss. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Linearity of expectation is the rule that expectation of a sum equals sum of expectations. It holds without independence. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. E[Σa_iX_i]=Σa_iE[X_i]. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Independence is unnecessary for means but crucial for many variance simplifications. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into portfolio return. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Variance is expected squared deviation from the mean. It measures dispersion and supports covariance algebra. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Var(X)=E[(X−μ)^2]=E[X²]−μ². Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Variance treats positive and negative deviations symmetrically. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into volatility. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Standard deviation is square root of variance. It restores the units of the underlying variable. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. σ=√Var(X). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. One standard deviation has no fixed tail meaning without a distributional model. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk reporting. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Covariance is expected product of joint deviations. It measures linear co-movement. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Cov(X,Y)=E[(X−μ_X)(Y−μ_Y)]. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Covariance magnitude depends on units. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into portfolio risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Correlation is standardised covariance. It enables comparison of linear dependence across scales. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. ρ=Cov(X,Y)/(σ_Xσ_Y). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Correlation does not specify the full joint distribution. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into dependence modelling. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Higher moments is moments beyond mean and variance such as skewness and kurtosis. They reveal asymmetry and tail properties hidden by volatility. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Standardised central moments define skewness and kurtosis. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Sample higher moments are noisy and sensitive to outliers. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into distribution diagnostics. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Skewness is a standardised third central moment. It describes asymmetry. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. γ_1=E[(X−μ)^3]/σ^3. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A single skewness number does not identify where all tail risk lies. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into return distributions. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Kurtosis is a standardised fourth central moment. It is related to tail weight and concentration. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. γ_2=E[(X−μ)^4]/σ^4, with excess kurtosis subtracting 3 under one convention. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Kurtosis is not synonymous with tail probability at every threshold. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into fat-tail analysis. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Bernoulli distribution is a two-outcome distribution such as default/nondefault over one horizon. It is the building block of many credit and operational models. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. P(X=1)=p, Var(X)=p(1−p). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Real default risk is conditional and heterogeneous; one p may be too crude. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into credit risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Binomial distribution is the distribution of successes across n independent identical Bernoulli trials. It models counts under strong assumptions. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. P(K=k)=C(n,k)p^k(1−p)^(n−k). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Loan defaults are not generally independent or identically distributed. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into default counts. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Poisson distribution is a count distribution often used for arrivals over time under a constant-rate independent-increment idealisation. It is useful for event-frequency modelling. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. P(N=k)=e^−λλ^k/k!. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Clustering, seasonality and overdispersion can make Poisson assumptions weak. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into operational events. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Normal distribution is a symmetric bell-shaped distribution determined by mean and variance. It supports analytic probability and portfolio formulas. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Z=(X−μ)/σ standardises to N(0,1). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Financial returns can exhibit skew, fat tails, jumps and volatility clustering. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into parametric VaR. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Standard normal distribution is the normal distribution with mean zero and variance one. It turns diverse normal variables into one probability table or CDF. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Z=(X−μ)/σ. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Using normal z-scores on visibly non-normal tails can understate risk. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into quantile calculations. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Lognormal distribution is a positive distribution generated when the logarithm of a variable is normal. It is used in simplified asset-price models because prices remain positive. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. If ln X is normal, X is lognormal. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Asset prices can jump and return volatility can vary over time. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into price modelling. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Student t distribution is a symmetric heavy-tailed family indexed by degrees of freedom. It can represent more tail mass than the normal distribution. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. As degrees of freedom grow, t approaches normal. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Symmetry still prevents modelling skew unless extended. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk returns. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Chi-square distribution is a positive distribution arising from sums of squared standard normals. It appears in variance estimation and statistical tests. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. If Z_i iid N(0,1), ΣZ_i² is chi-square. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Using asymptotic tests mechanically with dependent financial data can mislead. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into inference. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. F distribution is a distribution related to ratios of scaled chi-square variables. It appears in variance comparisons and regression testing. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. F statistics compare explained and residual variance structures. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Financial residual assumptions may violate textbook independence/homoskedasticity. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into model testing. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Exponential distribution is a continuous waiting-time distribution with memoryless property. It connects to simple event-arrival models. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. f(t)=λe^−λt for t≥0. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Credit and operational hazards often vary over time, violating constant intensity. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into hazard models. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Gamma distribution is a positive flexible distribution generalising waiting-time structures. It can model positive quantities and intensity heterogeneity. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Shape and scale determine mean and variance. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Fit convenience does not guarantee tail adequacy. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into credit and insurance models. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Beta distribution is a flexible distribution on [0,1]. It is natural for uncertain probabilities, proportions and recovery rates. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Parameters α,β shape mean and dispersion. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Bounded support may still be too simple for multimodal behaviour. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into Bayesian modelling. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Mixture distribution is a weighted combination of component distributions. It can represent regimes or heterogeneous populations. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. f(x)=Σπ_k f_k(x), Σπ_k=1. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Mixtures can fit data well while components remain hard to identify. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into regime risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Empirical distribution is the observed sample distribution without imposing a parametric family. It preserves realised shapes in the data. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Assign mass 1/n to each observation in the simplest empirical model. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. History may omit future regimes and rare events. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into historical simulation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Order statistic is a sorted sample value. It underlies empirical quantiles and nonparametric VaR. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. X_(k) denotes the kth ordered observation. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Small samples make tail order statistics unstable. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk estimation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Law of large numbers is the convergence principle that sample averages approach expectations under suitable conditions. It justifies many statistical estimates. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Sample mean converges to E[X] under standard assumptions. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Dependence, heavy tails and structural breaks can slow or invalidate simple intuition. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into estimation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Central limit theorem is a family of results under which suitably normalised sums can approach a normal distribution. It explains why aggregation can look normal in many settings. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. √n(mean−μ) converges in distribution under conditions. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. It does not say raw financial returns are normal, nor does it erase extreme-tail issues automatically. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into inference. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Sampling error is variation in an estimate caused by observing a finite sample. It is unavoidable even under a correct model. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Standard errors quantify estimator uncertainty under assumptions. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A precise-looking probability may have wide sampling uncertainty. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into model risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Confidence interval is a procedure producing intervals with stated long-run coverage under a statistical model. It communicates estimation uncertainty. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Estimate ± critical value×standard error in simple cases. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A 95% confidence interval is not generally a 95% posterior probability statement. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into inference. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Hypothesis test is a decision procedure comparing data with a null model. It can test calibration or model restrictions. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use a test statistic and reference distribution. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Failure to reject is not proof the model is true. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into backtesting. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. p-value is the probability, under the null model, of data at least as extreme as observed by the test statistic. It is evidence calibration, not the probability the null is true. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Compute from the test statistic’s null distribution. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Misinterpreting p as P(null|data) reverses conditioning. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into model validation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Type I error is rejecting a true null hypothesis. It is controlled by the significance level in classical testing. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. P(reject H0|H0 true)=α. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A small α raises Type II error unless sample size/power changes. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk model tests. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Type II error is failing to reject a false null. It captures missed-model-failure risk. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. β=P(fail reject H0|alternative true). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Reporting only significance without power hides detection weakness. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into backtesting. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Conditional expectation is the expected value given an information set or event. It is foundational to forecasting and pricing. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. E[X|Y] is itself a random variable in general. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Replacing conditional models with unconditional averages discards information. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into credit and market forecasts. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Probability tree is a branching representation of sequential events and conditional probabilities. It makes path dependence and updating visible. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Multiply branch probabilities along a path and add mutually exclusive paths. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Branches must represent mutually exclusive and collectively exhaustive outcomes at each node. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into scenario analysis. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Bayesian prior is a probability distribution representing uncertainty before new evidence. It provides the starting point for Bayesian updating. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Posterior∝Likelihood×Prior. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A prior should not be hidden as though it were data-free objectivity. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into Bayesian risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Likelihood is the probability model for observed data as a function of unknown parameters or states. It tells how compatible evidence is with candidate hypotheses. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. L(θ|data)∝P(data|θ). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Likelihood is not a probability distribution over θ until combined/normalised in a Bayesian framework. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into estimation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Posterior distribution is the updated probability distribution after applying Bayes’ rule. It combines prior information with observed evidence. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. p(θ|data)∝p(data|θ)p(θ). Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Posterior quality inherits model and prior misspecification. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into credit updating. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Base-rate fallacy is underweighting the prior prevalence of an event when interpreting a signal. It creates false certainty from noisy classifiers. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Posterior odds=prior odds×likelihood ratio. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. High test accuracy can still produce many false positives when the event is rare. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into fraud detection. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Loss distribution is the probability distribution of monetary or economic loss over a stated horizon. It is the central object behind VaR and ES. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Define L so larger positive values mean larger losses, then keep sign convention consistent. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Switching between return and loss conventions can invert quantiles. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk metrics. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Value at Risk is a quantile-based loss threshold for a chosen horizon and confidence level. It answers ‘how bad is the threshold exceeded only in a specified tail fraction under this model?’ The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. VaR_α=quantile_α(L) under the loss convention. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. VaR does not tell the average or maximum loss beyond the threshold. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into market risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Expected Shortfall is the average tail loss beyond a chosen quantile under standard continuous formulations. It adds severity information beyond VaR. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. ES_α=E[L|L≥VaR_α] in a simple continuous model. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Tail estimates can be noisy, model-sensitive and data-hungry. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into Basel market risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Subadditivity is the diversification-consistency property ρ(X+Y)≤ρ(X)+ρ(Y) for a risk measure. It is one reason Expected Shortfall is favoured in coherent-risk frameworks. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Check the inequality for the chosen risk measure and positions. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Some VaR setups can violate subadditivity. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk aggregation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Coherent risk measure is a risk-measure framework satisfying properties such as monotonicity, translation invariance, positive homogeneity and subadditivity. It formalises desirable aggregation behaviour. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Expected Shortfall satisfies coherence under standard definitions. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Coherence is not the same as statistical accuracy or model adequacy. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk theory. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Parametric VaR is VaR calculated from an assumed parametric distribution. It can be fast and transparent. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. For normal loss, VaR=μ_L+z_ασ_L. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Normality can understate skewed or heavy-tail risk. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk engines. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Historical VaR is VaR estimated from empirical historical P&L or returns. It avoids an explicit parametric distribution family. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Take the empirical loss quantile. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. The historical window may not contain the regimes that matter next. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into market risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Monte Carlo VaR is VaR estimated from simulated scenarios generated by a model. It can handle nonlinear portfolios and complex factors. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Simulate factors, revalue portfolio, sort losses, read quantile. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Output quality depends on model, calibration and simulation convergence. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into derivatives risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Monte Carlo simulation is repeated random sampling from a specified model to approximate output distributions. It converts difficult analytic problems into computational experiments. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Estimate E[g(X)] by the average of g(X_i) over simulated draws. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Simulation error can be tiny while model error remains huge. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into quantitative finance. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Pseudo-random number is a deterministic computer-generated sequence designed to mimic random draws. It powers reproducible simulation. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use seeded generators and tested algorithms. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Poor generators or accidental correlation can bias simulation. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into computational risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Variance reduction is techniques reducing Monte Carlo estimator noise without simply multiplying simulation count. It improves efficiency. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Antithetic variables, control variates and importance sampling are examples. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Bad controls or weights can increase instability. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into simulation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Bootstrap is resampling observed data to approximate estimator or scenario uncertainty. It is a data-driven simulation technique. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Sample observations with replacement in the basic bootstrap. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Naive resampling breaks serial dependence in time series. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into model validation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Block bootstrap is resampling blocks rather than individual observations. It preserves some time dependence. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Choose block length to balance dependence and sample variety. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Block-length choice is itself a modelling decision. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into financial time series. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Scenario analysis is evaluating outcomes under explicitly chosen states rather than relying only on estimated distribution frequencies. It is essential for risks not well represented by historical samples. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Map each scenario to factor shocks and revalue. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A severe scenario need not have a known probability. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into stress testing. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Stress testing is evaluation under extreme but plausible or exploratory adverse conditions. It probes model failure and capital/liquidity resilience. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use multi-factor shocks, behavioural responses and full repricing. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Stress tests can be gamed by choosing convenient scenarios. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into bank risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Reverse stress test is starting from an unacceptable outcome and asking which scenarios could cause it. It reveals hidden fragility. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Solve or search for factor combinations breaching a threshold. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. It identifies pathways, not necessarily their probability. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into resilience. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Backtesting is comparing risk forecasts with realised outcomes. It tests calibration and operational use. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Count exceptions, examine independence and study residuals. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Passing a weak test does not prove the model is correct. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into VaR validation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Exception is an observed loss exceeding a VaR forecast threshold. Its frequency is central to VaR coverage tests. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Compare observed exception count with expected count under calibration. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Clustering of exceptions can signal volatility dynamics even if total count looks acceptable. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk monitoring. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Kupiec coverage test is a likelihood-ratio test for whether VaR exception frequency matches the nominal rate. It checks unconditional coverage. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Compare observed and expected exception proportions. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. It does not test independence or severity beyond VaR. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into VaR backtesting. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Christoffersen independence test is a test for clustering or dependence in VaR exceptions. It checks whether breaches occur independently over time under the model assumptions. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use transition counts between exception/non-exception states. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Small samples can give low power. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into risk validation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Calibration is agreement between predicted probabilities/distributions and realised frequencies over appropriate groups or horizons. It asks whether 10% predictions happen about 10% of the time in the relevant sense. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use reliability curves, PIT tests or event-frequency comparisons. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A model can be calibrated but have weak ranking power. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into forecast quality. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Discrimination is the ability to rank higher-risk cases above lower-risk cases. It is different from calibration. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. AUC/ROC is one common binary-ranking measure. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A well-ranked model can still output badly calibrated probabilities. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into credit scoring. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Brier score is mean squared error of probability forecasts for binary outcomes. It is a proper scoring rule sensitive to calibration and resolution. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Brier=mean(p_i−y_i)^2. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Class imbalance and decomposition matter for interpretation. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into probability forecasts. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Log loss is negative log-likelihood score for probabilistic classification. It strongly penalises confident wrong probabilities. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. −[y ln p+(1−y)ln(1−p)]. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Predictions of exactly 0 or 1 create infinite penalty when wrong. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into model evaluation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Probability integral transform is mapping observations through the forecast CDF. It can test full-distribution calibration. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. u_t=F_t(x_t) should be uniform under a correct continuous conditional distribution. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Dependence and parameter estimation complicate tests. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into distribution validation. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Extreme value theory is statistical modelling focused directly on extreme tails. It can be more appropriate than fitting one distribution to the entire range. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Peaks-over-threshold uses a Generalised Pareto approximation under conditions. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Threshold choice and regime stability are major model risks. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into tail risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Generalised Pareto distribution is the asymptotic tail model used in peaks-over-threshold EVT under conditions. It parameterises tail scale and heaviness. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Exceedances above a high threshold are modelled with GPD. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Too-low thresholds bias tail shape; too-high thresholds leave too little data. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into rare losses. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Tail dependence is dependence of extreme outcomes across variables. It matters when diversification fails precisely in stress. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Tail-dependence coefficients study joint extreme probabilities. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Linear correlation can be modest while tail dependence is material. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into systemic risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Copula is a function joining marginal distributions into a multivariate joint distribution. It lets dependence be modelled separately from marginals. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. C(u,v) links F_X and F_Y under Sklar’s theorem. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A copula calibrated to normal periods may fail in crises. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into portfolio and credit risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Gaussian copula is a copula with dependence inherited from a multivariate normal latent structure. It is analytically convenient and historically influential. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Transform marginals to latent Gaussian variables with correlation matrix. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. It has weak tail dependence relative to many crisis patterns. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into credit portfolio modelling. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. t copula is a heavy-tail copula with symmetric tail dependence. It can represent more joint extremes than a Gaussian copula. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Dependence determined by correlation matrix and degrees of freedom. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Symmetric tails may still be inappropriate. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into multi-asset risk. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Model risk is the possibility that the chosen probability model, parameters or implementation is wrong for the decision. It is unavoidable in risk mathematics. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use challenger models, stress tests, validation and governance. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. More decimal places do not reduce structural model error. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into all risk systems. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Parameter risk is uncertainty in estimated model parameters. It should be distinguished from randomness conditional on parameters. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use confidence/posterior distributions and sensitivity analysis. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Plug-in estimates can understate uncertainty. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into forecasting. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Regime risk is the possibility that the data-generating process changes. It attacks stationarity assumptions. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use rolling estimates, change-point models and stress regimes. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. A model calibrated to one regime can fail abruptly in another. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into financial time series. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Liquidity risk is loss or constraint arising because positions cannot be traded near model value quickly enough. It is not captured by a static return distribution alone. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Add liquidation horizon, bid-ask, depth and price impact. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Daily VaR can look small for an asset that cannot be sold during stress. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into banking and markets. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Wrong-way risk is dependence in which exposure increases when counterparty credit quality deteriorates. It is a joint-distribution problem. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Model exposure and default together rather than independently. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Independence assumptions can materially understate counterparty loss. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into counterparty credit. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. Systemic risk is risk emerging from interactions across institutions, markets and feedback loops. It is fundamentally multivariate and networked. The first discipline is to state the horizon and information set, because the same event can have different probabilities under different conditions. Mathematics. Use joint scenarios, network exposure matrices and contagion mechanisms. Keep return and loss sign conventions explicit. A left-tail return event becomes a right-tail loss event after changing sign, and a quantile changes interpretation accordingly. Failure mode. Summing standalone probabilities misses feedback and common shocks. Jo’s diagnostic is to ask what assumption allows the formula to be used. If independence, normality, stationarity or identical distribution is required, write it rather than letting it disappear inside the notation. Connection. The concept feeds into financial stability. Ryan would verify it with either an exact probability calculation, a simulation cross-check or an empirical calibration test. This makes probability an auditable model rather than a decorative percentage. A loan has one-year probability of default 2%, exposure at default S$100,000 and loss given default 40% in a simplified deterministic-LGD model. Expected loss =PD×LGD×EAD=0.02×0.40×100,000=S$800. The expected loss is not the most likely realised loss. In the simplest two-state model, the actual credit loss might be S$0 with 98% probability or S$40,000 with 2% probability. S$800 is the probability-weighted average across repeated comparable exposures or model repetitions. This distinction between expectation and realised outcome is foundational in finance. Suppose 1% of transactions are truly fraudulent. A detector flags 90% of frauds but also flags 5% of legitimate transactions. Among 10,000 transactions, expect 100 frauds, 90 flagged; and 9,900 legitimate transactions, 495 flagged. Total flags=585. Probability a flagged transaction is truly fraud is 90/585≈15.38%, not 90%. The 90% number was sensitivity P(flag|fraud); the decision needs P(fraud|flag). Base rates matter. This is one of the most important probability reversals in underwriting, medical testing, fraud monitoring and machine learning. Assume a one-day portfolio P&L is normal with mean zero and standard deviation S$1 million. A 99% one-tailed normal quantile is about 2.326 standard deviations. Parametric 99% one-day VaR is therefore about S$2.326 million under the model. The statement means the loss threshold is exceeded with model probability about 1% over the chosen one-day horizon. It does not mean the maximum loss is S$2.326 million. The entire result rests on the normal distribution, stable volatility and the mapping from factor moves to P&L. For a standard normal loss variable, expected shortfall beyond quantile z_α has a closed-form ratio involving the normal density φ(z_α)/(1−α). Multiply by portfolio standard deviation and add the mean under the chosen loss convention. At 97.5% confidence, ES exceeds the corresponding VaR because it averages losses deeper in the tail rather than stopping at the threshold. The Basel market-risk framework uses a 97.5% one-tailed expected-shortfall framework with additional regulatory specifications, including liquidity-horizon treatment. A classroom Gaussian ES formula is therefore a foundation, not the regulatory capital formula. Take 1,000 historical daily portfolio losses, sort them from smallest to largest and inspect the upper tail. A 99% empirical VaR is associated with roughly the 10 worst observations, with exact index convention needing to be stated. This method preserves historical asymmetry and fat tails in the sample but assumes the past window is informative about the future. If the sample contains no crisis comparable to the next one, the empirical tail cannot invent it. That is why historical simulation and stress testing complement rather than replace each other. Suppose the quantity of interest is an expected discounted payoff. Simulate N independent scenarios X_i and estimate the expectation with the sample mean of g(X_i). Under standard finite-variance conditions, Monte Carlo standard error shrinks approximately as 1/√N. To cut simulation noise in half, roughly four times as many independent draws are needed. To cut it to one tenth, roughly one hundred times as many draws are needed. This square-root law explains why variance reduction is economically valuable. But simulation convergence only addresses numerical error conditional on the model. A perfectly converged wrong model remains wrong. Imagine two joint models with the same marginal return distributions and the same linear correlation 0.3. One uses a Gaussian copula; another uses a low-degrees-of-freedom t copula. Their everyday co-movement can look similar, yet the t-copula model can assign materially more probability to joint extremes. A portfolio risk report based only on correlation can therefore miss the difference. Tail dependence asks a more specific question: when one variable is extreme, how likely is the other to be extreme too? This matters in credit portfolios, contagion analysis and diversified portfolios during stress. A borrower can enter a normal economy with probability 70% or recession with probability 30%. Conditional one-year default probabilities are 1% and 6% respectively. By the law of total probability, unconditional PD=0.70×0.01+0.30×0.06=2.5%. Now change the recession probability to 50% while keeping conditional PDs unchanged. Unconditional PD becomes 3.5%. The risk changes because scenario weights changed even though within-scenario borrower behaviour did not. This simple tree is the conceptual ancestor of multi-scenario expected-credit-loss systems. If 100 loans each independently default with probability 2%, a binomial model gives expected defaults np=2 and variance np(1−p)=1.96. This is mathematically straightforward. But real borrowers share macroeconomic drivers. Positive default correlation increases the probability of clustered losses relative to the independent binomial model. The expected count can remain 2 while the tail becomes much heavier. This is a powerful example of why expected loss and unexpected loss are different objects. Suppose a bank asks what happens if interest rates rise 200bp, unemployment rises sharply and property values fall 25%. The scenario can be evaluated even if management does not claim it has a precise 0.7% probability. Scenario severity and scenario probability are separate dimensions. A stress test is valuable because some important states are too rare, structurally changing or model-uncertain for a trustworthy frequency estimate. The mathematical discipline is to state whether a scenario is probabilistic, historical, hypothetical or exploratory. Expected loss asks for the average. Standard deviation asks about dispersion. VaR asks for a quantile threshold. Expected Shortfall asks about average loss beyond a threshold. Maximum drawdown asks about path-dependent peak-to-trough decline. Stress loss asks what happens under a chosen adverse scenario. None subsumes all the others. This is why risk systems need a dashboard of complementary measures rather than one universal number. The appropriate metric depends on the decision: pricing, limit setting, capital, liquidity, resilience, underwriting or communication. Mira’s rule is simple: before comparing two risk numbers, ask whether they answer the same question. The Basel Framework defines market risk as risk of losses from movements in market prices for specified banking/trading-book exposures. Its current framework uses a standardised approach and an internal-models architecture with expected shortfall in the latter, plus explicit treatment of liquidity horizons and non-modellable risk factors. This page does not reproduce a bank’s regulatory calculation. The purpose is to show the probability mathematics underneath concepts such as quantiles, tail averages, dependence and simulation. Regulatory capital adds prescribed definitions, classifications, horizons, correlations, stress periods, model-eligibility rules and supervisory governance. The distinction protects against a common educational error: learning a generic formula and assuming it is the regulatory implementation. A probability forecast should be judged against outcomes. If a credit model assigns 10% PD to many similar cases, roughly 10% should default over the relevant horizon in a well-calibrated stable setting, subject to sampling noise and conditioning. If the realised rate is systematically 2% or 25%, the probabilities need investigation. Calibration is not the same as discrimination. A model can rank risky borrowers correctly but output probabilities that are too high or too low. Another model can be calibrated on average but fail to distinguish safe from risky borrowers. Both dimensions matter. This is why Brier score, log loss, ROC/AUC and calibration curves answer different validation questions. A model becomes more trustworthy when the team knows what evidence would count against it. For a VaR model, excessive or clustered exceptions are warning signals. For a probability model, poor calibration by segment matters. For a distribution forecast, non-uniform probability-integral transforms can reveal shape errors. For a correlation model, stress co-movement beyond model ranges can reveal dependence failure. This is a general scientific habit: specify the test before the failure becomes politically inconvenient. Risk mathematics should make it easier to discover error, not easier to defend a number. Ethan treats every metric as provisional until its monitoring loop is defined. P(signal|bad) is not P(bad|signal). Use Bayes and base rates. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Expected loss is an average, not the guaranteed realised loss. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Gaussian formulas can materially understate skewed or fat-tailed risks. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Nonlinear and tail dependence can remain. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. VaR is a quantile threshold, not a cap. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. ES is a tail average, not the maximum conceivable loss. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Past windows cannot contain regimes that never occurred in the sample. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. More paths reduce Monte Carlo noise, not structural model error. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Plug-in probabilities and volatilities can look more certain than the evidence allows. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. One-day, ten-day and one-year risks are not directly comparable without a model. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Left-tail return quantiles and right-tail loss quantiles reverse interpretation. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Exception clustering and severity can reveal failures hidden by total counts. The repair is to restate the random variable, conditioning information, horizon and probability model, then recompute the measure from the distribution rather than from memory. Situation. A lender estimates one-year default probability for a borrower under changing employment and interest-rate conditions. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Build conditional PDs by scenario, combine them with scenario weights and update as new information arrives. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. The result is model-conditional, not a statement that one borrower will default fractionally. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. A rare fraud base rate meets a classifier with high sensitivity but nontrivial false-positive rate. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Use Bayes’ rule to calculate posterior fraud probability after a flag. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Ignoring the base rate can overwhelm operations with false alerts. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. The desk needs a distribution of one-day losses from market moves. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Choose factor distributions and dependence, revalue positions, then calculate VaR, ES and stress losses. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. A one-day model does not automatically describe liquidation loss over longer stressed horizons. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. The last 500 daily factor moves are replayed through today’s portfolio. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Revalue today’s positions under each historical move and sort the P&L distribution. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. The method preserves observed joint moves but cannot represent unseen scenarios without augmentation. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. Nonlinear derivatives make closed-form portfolio loss distributions difficult. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Simulate underlying factors, reprice each instrument and aggregate scenario P&L. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Numerical convergence should be separated from volatility-model and correlation-model risk. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. PDs and LGDs both rise under recession conditions. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Use a joint macro scenario to move default frequency and recovery severity together. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Stressing only PD while holding LGD benign can miss wrong-way macro dependence. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. Market positions can be valued but cannot be sold near model prices quickly. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Add liquidation horizon, bid-ask widening and price impact to the stress scenario. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Market-price distributions alone do not capture execution constraints. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. A 99% VaR model produces 15 exceptions in 250 days. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Compare observed exception frequency with the nominal expectation and test clustering. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Exception count alone does not assess tail severity or all forms of model misspecification. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. Normal, historical and t-distribution models produce similar volatility but different Expected Shortfall. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Compare tail quantiles and average exceedance losses, then examine which assumptions drive differences. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Model dispersion is itself information about tail uncertainty. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. Two asset classes have modest historical correlation but share a funding-liquidity channel. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Build a scenario with stronger joint downside dependence and compare losses with the Gaussian-correlation model. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Correlation is not a complete map of crisis dependence. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. A borrower’s prior risk is updated after new financial information. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Specify prior PD, likelihood of the evidence under default/nondefault states and calculate the posterior. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. The update is only as credible as the likelihood model and evidence quality. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Situation. Management asks what combination of default, market and liquidity shocks would breach capital tolerance. The mathematical objective is to define the random object and information set before selecting a risk metric. Method. Start at the breach condition and search scenario space for plausible pathways. Adrian defines the events and horizon, Jo checks conditioning, Aisha checks distribution and units, and Ryan compares analytic or simulated results with an independent calculation. Boundary. Reverse stress identifies vulnerabilities without pretending every pathway has a precise probability. Mira then asks what evidence would falsify or recalibrate the model. That turns the risk number into a monitored hypothesis rather than a permanent fact. Probability is not certainty with decimals. It is a disciplined language for uncertainty. Expected value, variance, quantiles and tail averages are different projections of a distribution. Conditional probability and Bayes’ rule make information explicit. Simulation turns a model into a scenario engine. Backtesting and calibration force the forecasts to meet reality. The strongest financial risk mathematics therefore keeps four layers separate: the uncertain world, the probability model, the summary risk metric and the decision made from that metric. Confusing the layers is one of the fastest ways to make a technically correct formula operationally wrong. This foundation now supports the later Banking And Finance Mathematics lanes for credit risk, market risk, bank stress testing, derivatives, capital and systemic risk without competing with their specialist mechanisms. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Hold marginal loss distributions fixed and vary the dependence model or correlation. Compare portfolio variance, VaR, ES and joint-exceedance frequency. The exercise separates marginal risk from interaction risk. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Run a Monte Carlo estimate with 10,000, 40,000 and 160,000 paths to observe sampling convergence. Then change the underlying distributional assumption. The first comparison measures numerical error; the second exposes structural model sensitivity. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Create three macro states with probabilities summing to one and assign conditional default or loss probabilities to each. Calculate the unconditional probability by the law of total probability. Then change only the scenario weights and explain why the unconditional risk moves. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Choose a rare event and a noisy signal. Compute P(signal|event), P(signal|not event), the base rate and then P(event|signal) using Bayes. Repeat at three different base rates to see how prevalence changes posterior interpretation. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution. Construct or simulate a loss sample with the same mean and standard deviation under two different tail shapes. Calculate VaR and Expected Shortfall at several confidence levels. Observe where volatility fails to reveal tail differences. Complete the lab with a written statement of the random variable, horizon, sign convention and information set. Ben should be able to reproduce the arithmetic; Clara should identify which input came from data versus assumption; Ethan should specify a monitoring statistic or backtest that would trigger review. Finish by changing one assumption that is normally hidden—independence, normality, stationarity, constant volatility or fixed correlation—and record which result changes most. The purpose is to build sensitivity to model structure, not merely speed at formula substitution.14. Probability mass function
15. Probability density function
16. Cumulative distribution function
17. Quantile
18. Expected value
19. Linearity of expectation
20. Variance
21. Standard deviation
22. Covariance
23. Correlation
24. Higher moments
25. Skewness
26. Kurtosis
27. Bernoulli distribution
28. Binomial distribution
29. Poisson distribution
30. Normal distribution
31. Standard normal distribution
32. Lognormal distribution
33. Student t distribution
34. Chi-square distribution
35. F distribution
36. Exponential distribution
37. Gamma distribution
38. Beta distribution
39. Mixture distribution
40. Empirical distribution
41. Order statistic
42. Law of large numbers
43. Central limit theorem
44. Sampling error
45. Confidence interval
46. Hypothesis test
47. p-value
48. Type I error
49. Type II error
50. Conditional expectation
51. Probability tree
52. Bayesian prior
53. Likelihood
54. Posterior distribution
55. Base-rate fallacy
56. Loss distribution
57. Value at Risk
58. Expected Shortfall
59. Subadditivity
60. Coherent risk measure
61. Parametric VaR
62. Historical VaR
63. Monte Carlo VaR
64. Monte Carlo simulation
65. Pseudo-random number
66. Variance reduction
67. Bootstrap
68. Block bootstrap
69. Scenario analysis
70. Stress testing
71. Reverse stress test
72. Backtesting
73. Exception
74. Kupiec coverage test
75. Christoffersen independence test
76. Calibration
77. Discrimination
78. Brier score
79. Log loss
80. Probability integral transform
81. Extreme value theory
82. Generalised Pareto distribution
83. Tail dependence
84. Copula
85. Gaussian copula
86. t copula
87. Model risk
88. Parameter risk
89. Regime risk
90. Liquidity risk
91. Wrong-way risk
92. Systemic risk
Worked Example 1: Expected Loss
Worked Example 2: Bayes and a Rare Event
Worked Example 3: Normal VaR
Worked Example 4: Expected Shortfall Under a Normal Model
Worked Example 5: Historical VaR
Worked Example 6: Monte Carlo Convergence
Worked Example 7: Correlation Is Not Tail Dependence
Worked Example 8: Probability Tree for a Loan
Worked Example 9: Binomial Default Count and Dependence Warning
Worked Example 10: Stress Without Probability
Risk Metrics Are Different Questions
The Basel Market-Risk Boundary
Probability Models Need Calibration
Risk Models Need Falsifiers
A Professional Probability-and-Risk Workflow
Common Failure Modes
1. Probability reversal
2. Expectation as prediction
3. Normality by convenience
4. Zero correlation as independence
5. VaR as maximum loss
6. Expected Shortfall as worst-case loss
7. Historical sample as complete future
8. Simulation count as model quality
9. Ignoring parameter uncertainty
10. Mixing horizons
11. Mixing return and loss signs
12. Backtesting only frequency
Formula Map
Concept Formula Meaning Conditional probability P(A|B)=P(A∩B)/P(B) Probability after conditioning on B. Bayes P(A|B)=P(B|A)P(A)/P(B) Update prior probability with evidence. Expected value E[X]=Σxp(x) or ∫xf(x)dx Probability-weighted average. Variance E[(X−μ)²] Squared dispersion around mean. Covariance E[(X−μ_X)(Y−μ_Y)] Joint linear co-movement. Correlation Cov(X,Y)/(σ_Xσ_Y) Standardised covariance. VaR Quantile of loss L Tail threshold at chosen confidence/horizon. Expected Shortfall E[L|L≥VaR_α] in simple continuous form Average loss beyond the VaR threshold. Monte Carlo mean (1/N)Σg(X_i) Simulation estimator of an expectation. Authoritative Reference Map
Connected Banking And Finance Mathematics Route
Applied Case Study 1: A mortgage default probability
Applied Case Study 2: Credit-card fraud screening
Applied Case Study 3: A daily trading-book loss model
Applied Case Study 4: A historical-simulation engine
Applied Case Study 5: A Monte Carlo options portfolio
Applied Case Study 6: A loan portfolio recession stress
Applied Case Study 7: A liquidity shock
Applied Case Study 8: A VaR backtest
Applied Case Study 9: An ES model comparison
Applied Case Study 10: A dependence stress
Applied Case Study 11: A Bayesian credit update
Applied Case Study 12: A reverse stress test
Final Principle
Deep Practice Lab 1: Build a probability tree
Deep Practice Lab 2: Reverse a conditional probability
Deep Practice Lab 3: Compare tail metrics
Deep Practice Lab 4: Stress dependence
Deep Practice Lab 5: Separate numerical and model error
Deep Practice Lab 6: Build a probability tree
Deep Practice Lab 7: Reverse a conditional probability
Deep Practice Lab 8: Compare tail metrics
Deep Practice Lab 9: Stress dependence
Deep Practice Lab 10: Separate numerical and model error
Deep Practice Lab 11: Build a probability tree
Deep Practice Lab 12: Reverse a conditional probability
Deep Practice Lab 13: Compare tail metrics
Deep Practice Lab 14: Stress dependence
Deep Practice Lab 15: Separate numerical and model error
Deep Practice Lab 16: Build a probability tree
Deep Practice Lab 17: Reverse a conditional probability
Deep Practice Lab 18: Compare tail metrics
Deep Practice Lab 19: Stress dependence
Deep Practice Lab 20: Separate numerical and model error
Deep Practice Lab 21: Build a probability tree
Deep Practice Lab 22: Reverse a conditional probability
Deep Practice Lab 23: Compare tail metrics
Deep Practice Lab 24: Stress dependence
Deep Practice Lab 25: Separate numerical and model error
Deep Practice Lab 26: Build a probability tree
Deep Practice Lab 27: Reverse a conditional probability
Deep Practice Lab 28: Compare tail metrics
Deep Practice Lab 29: Stress dependence
Deep Practice Lab 30: Separate numerical and model error
Deep Practice Lab 31: Build a probability tree
Deep Practice Lab 32: Reverse a conditional probability
Deep Practice Lab 33: Compare tail metrics
Deep Practice Lab 34: Stress dependence
Deep Practice Lab 35: Separate numerical and model error
Deep Practice Lab 36: Build a probability tree
Deep Practice Lab 37: Reverse a conditional probability
Deep Practice Lab 38: Compare tail metrics
Deep Practice Lab 39: Stress dependence
Deep Practice Lab 40: Separate numerical and model error
Deep Practice Lab 41: Build a probability tree
Deep Practice Lab 42: Reverse a conditional probability
Deep Practice Lab 43: Compare tail metrics
Deep Practice Lab 44: Stress dependence
Deep Practice Lab 45: Separate numerical and model error
Deep Practice Lab 46: Build a probability tree
Deep Practice Lab 47: Reverse a conditional probability
Deep Practice Lab 48: Compare tail metrics
Deep Practice Lab 49: Stress dependence
Deep Practice Lab 50: Separate numerical and model error
Deep Practice Lab 51: Build a probability tree
Deep Practice Lab 52: Reverse a conditional probability
Deep Practice Lab 53: Compare tail metrics