Credit risk mathematics turns one vague question—“what if the borrower does not pay?”—into a structured loss system built from probability of default (PD), loss given default (LGD), exposure at default (EAD), recovery, maturity, discounting, dependence and scenario assumptions. This long-form guide develops expected credit loss, unexpected credit loss, default probability, recovery mathematics, collateral, credit conversion factors, revolving exposures, ratings, scorecards, migration, concentration, IFRS 9 expected credit losses and Basel credit-risk concepts from first principles.

For readers searching for credit risk mathematics, PD LGD EAD, expected loss formula, probability of default, loss given default, exposure at default, recovery rate, expected credit loss, IFRS 9 ECL, Basel credit risk, IRB credit risk, credit scoring, default probability, credit migration, credit concentration, lifetime PD or loan-loss mathematics, the central identity is simple but powerful: Expected Loss ≈ PD × LGD × EAD in the simplest one-period model. The difficult work lies in defining each term correctly, matching the horizon, modelling dependence and understanding which framework the number belongs to.

Current authoritative frameworks preserve these distinctions. The Basel Framework uses PD, LGD, EAD and maturity as credit-risk components under the internal ratings-based architecture and treats expected loss separately from unexpected-loss capital. IFRS 9 defines expected credit losses as probability-weighted present values of contractual cash shortfalls and distinguishes 12-month from lifetime expected losses according to credit deterioration. CFA Institute’s 2026 credit-analysis material likewise treats exposure, default probability and recovery/LGD as core credit-modelling inputs. This page is educational mathematics, not lending, accounting, legal or investment advice.

50-Second Router

PD: probability that a defined default event occurs over a stated horizon.
LGD: percentage of exposure economically lost if default occurs, after recoveries and costs under the chosen definition.
EAD: exposure expected to be outstanding when default occurs; for revolving facilities it can exceed today’s drawn balance.
Expected loss: in a simple one-period model, PD × LGD × EAD.
Recovery rate: the complement of LGD only under compatible definitions; timing and costs matter.
Unexpected loss: adverse deviation around expected loss; this is why capital and loss provisions are not the same concept.
IFRS 9: ECL is a probability-weighted present value of cash shortfalls; Stage 1 and lifetime ECL should not be reduced to one crude PD×LGD×EAD multiplication without matching horizon and discounting.
Basel IRB: PD, LGD, EAD and maturity feed prescribed regulatory risk-weight functions, with floors and definitions.
Portfolio credit risk: expected loss adds linearly, but tail loss depends strongly on default dependence, concentration and recovery dependence.
Verification: calibrate PD, test LGD recoveries, validate EAD drawdown behaviour, reconcile expected loss and monitor realised outcomes.

The Central Proposition: Credit Loss Is Probability × Severity × Exposure

A credit loss needs three questions answered. First: will default happen? That is the domain of PD. Second: if default happens, how much of the exposure is economically lost? That is LGD. Third: how large will the exposure be at the moment of default? That is EAD. The one-period expected-loss identity multiplies these dimensions because it combines event probability, conditional severity and amount at risk.

The identity is useful precisely because the three components fail in different ways. PD can be wrong because the score is miscalibrated or the economy changes. LGD can be wrong because collateral values, recovery timing or workout costs differ from assumptions. EAD can be wrong because customers draw additional credit before default. A model that simply labels everything “credit risk” makes diagnosis harder.

Adrian’s discipline is to refuse a credit-risk number until its horizon and definition are visible. “PD is 2%” is incomplete. Is it a one-year through-the-cycle PD, a point-in-time 12-month PD, a cumulative lifetime PD, or a risk-neutral default probability inferred from market prices? The same borrower can have several legitimate numbers because the questions differ.

1. Credit risk

Credit risk is the possibility of economic loss because a borrower, issuer, counterparty or contractual obligor fails to meet agreed obligations or deteriorates in credit quality. It includes default loss but can also affect value before default through spread widening and migration. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. A simple one-period loss variable is L=EAD×LGD×I(default). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Treating every fall in bond price as realised default loss confuses market-value credit risk with default cash loss. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into loans, bonds and counterparties. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

2. Default event

Default event is the precisely defined state that triggers the model’s default indicator. The event definition determines what historical observations count as defaults. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. I(default)=1 when the governed default definition is met, otherwise 0. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Changing default definitions across data periods makes PD estimates incomparable. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into PD calibration. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

3. Probability of default

Probability of default is the probability that a defined default event occurs over a stated future horizon, conditional on the information set. It converts borrower uncertainty into a probabilistic input. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PD=P(default during horizon | information today). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A default rate observed in a sample is an ex post frequency; it is not automatically the forward PD for every borrower. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit scoring and IRB. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

4. One-year PD

One-year PD is a probability of default over the next year under the chosen methodology. Basel IRB uses one-year PD concepts for relevant exposures within its framework. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PD_1y=P(τ≤1 year | alive today). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Multiplying one-year PD by five is not generally a valid five-year cumulative PD. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into regulatory capital. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

5. Lifetime PD

Lifetime PD is the cumulative probability of default over the remaining life or a defined multi-period horizon. It is important for lifetime ECL and long-dated credit valuation. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. CumPD(T)=1−Survival(T). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A lifetime PD must account for survival to each later period; summing marginal PDs without conditioning can exceed one. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into IFRS 9 and credit curves. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

6. Marginal PD

Marginal PD is the unconditional probability that default occurs during one interval. It allocates cumulative default risk across time. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. mPD_t=P(t−1<τ≤t). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using cumulative PD directly as a period-specific default increment double counts earlier defaults. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into term structures. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

7. Conditional PD

Conditional PD is the probability of default in the next interval conditional on survival to its start. It is the natural period-by-period hazard-style quantity. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. q_t=P(default in t | survived through t−1). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Conditional and marginal PD are different when survival is less than one. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into hazard models. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

8. Survival probability

Survival probability is the probability that no default has occurred through a horizon. It complements cumulative default probability. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. S(T)=1−CumPD(T). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using 1−annual PD repeatedly without defining time-varying hazards assumes stable independent conditional structure. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit curves. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

9. Hazard rate

Hazard rate is an instantaneous or interval default intensity conditional on survival. It provides a time-local default representation. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Under constant continuous hazard λ, S(t)=e^(−λt). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Hazard intensity is not numerically identical to cumulative PD except in small-probability approximations. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into reduced-form credit models. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

10. Loss given default

Loss given default is the proportion of EAD economically lost if default occurs under a specified recovery and discounting methodology. It transforms a default event into monetary severity. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. LGD=(EAD−PV(net recoveries))/EAD in a simplified workout model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Ignoring recovery timing and workout costs can materially understate economic LGD. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into recovery modelling. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

11. Recovery rate

Recovery rate is the fraction of exposure recovered after default under the chosen definition. It is closely linked to LGD but only under consistent timing and cost assumptions. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Recovery≈1−LGD in the simplest zero-cost contemporaneous model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Undiscounted gross recovery can be high while economic LGD remains high because recovery is delayed and costly. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into workout analysis. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

12. Discounted recovery

Discounted recovery is the present value of recovery cash flows after default. It recognises that S$1 recovered today is worth more than S$1 recovered years later. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PV(R)=ΣR_t/(1+r)^t. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Comparing nominal recovery percentages across cases with very different workout times can be misleading. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into LGD estimation. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

13. Workout cost

Workout cost is direct and indirect costs incurred in collections, restructuring, legal enforcement or collateral realisation. Costs reduce net recovery. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. NetRecovery=GrossRecovery−WorkoutCosts. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Ignoring costs can bias LGD downward, especially for small exposures or complex collateral. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into collections and collateral. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

14. Downturn LGD

Downturn LGD is an LGD concept reflecting loss severity under adverse economic conditions. Recoveries can deteriorate precisely when defaults rise. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Model LGD conditional on stressed macro/collateral conditions. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using benign average recoveries for stress capital can underestimate joint PD-LGD deterioration. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into Basel capital. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

15. Exposure at default

Exposure at default is the amount economically exposed when default occurs. It can differ from today’s balance because facilities amortise, draw, accrue or change before default. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. EAD=drawn balance + expected conversion of eligible undrawn commitments in a simplified model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using today’s outstanding amount for a revolving credit line can understate future default exposure. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit cards and commitments. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

16. Credit conversion factor

Credit conversion factor is a factor converting undrawn commitments into expected or regulatory exposure. It maps contingent commitments into an EAD contribution. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. EAD≈Drawn+CCF×Undrawn under a simplified static formula. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. CCF is not a universal behavioural constant and may rise as borrower quality deteriorates. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into off-balance-sheet exposure. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

17. Revolving exposure

Revolving exposure is a facility in which the borrower can draw, repay and redraw within limits. Its EAD is dynamic and behaviourally linked to credit deterioration. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. EAD at default depends on utilisation path before τ. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Borrowers often draw more before default, creating adverse exposure dynamics. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit cards. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

18. Term-loan EAD

Term-loan EAD is the default exposure for a scheduled-amortisation facility. It typically declines through contractual principal payments unless arrears, capitalised interest or further advances change it. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. EAD_t can be projected from the amortisation schedule and default timing. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using original principal as EAD throughout the life overstates later exposure. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into loan mathematics. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

19. Expected loss

Expected loss is the probability-weighted average credit loss over the stated horizon under the chosen model. It is the centre of the loss distribution, not the worst case. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. EL=PD×LGD×EAD in the simplest one-period deterministic-LGD/EAD model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Expected loss is not regulatory capital and not a maximum-loss forecast. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into pricing and provisioning. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

20. Unexpected loss

Unexpected loss is the adverse variation around expected loss. It motivates capital and tail-risk analysis beyond average loss. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. UL can be represented by variance, quantile loss minus EL, or prescribed capital formulas depending on framework. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Subtracting EL from VaR and calling the remainder universal capital ignores framework definitions. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into capital. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

21. Credit loss distribution

Credit loss distribution is the probability distribution of portfolio losses over a horizon. It contains expected loss, variance and tail losses together. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. L=Σ_i EAD_i LGD_i I_i(default). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Expected loss alone can stay unchanged while tail loss explodes because defaults become correlated. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into portfolio credit risk. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

22. Default indicator

Default indicator is a Bernoulli variable equal to one on default and zero otherwise. It provides the simplest mathematical representation of borrower state. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. D_i∈{0,1}, E[D_i]=PD_i. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Binary default ignores migration and pre-default credit deterioration. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into portfolio models. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

23. Expected portfolio loss

Expected portfolio loss is the sum of expected losses across exposures under linear expectation. Dependence does not change the arithmetic sum of individual expected values. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. E[L]=ΣPD_iLGD_iEAD_i under deterministic LGD/EAD. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Low expected loss does not imply low portfolio tail loss. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into portfolio aggregation. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

24. Default correlation

Default correlation is dependence between borrower default events. It determines clustering of losses and portfolio tail thickness. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Corr(D_i,D_j) depends on joint default probability. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Assuming independence can make a concentrated portfolio look falsely diversified. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit concentration. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

25. Asset correlation

Asset correlation is correlation of latent creditworthiness variables in structural or regulatory factor models. It is one way to generate correlated defaults. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Z_i=√ρY+√(1−ρ)ε_i in a one-factor teaching model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Latent asset correlation is not the same as observed equity-return correlation. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into Basel-style factor models. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

26. Concentration risk

Concentration risk is risk from large single-name, sector, geography or common-factor exposures. It increases loss variability beyond granular-portfolio assumptions. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. HHI=Σs_i² is one simple exposure-concentration measure. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A thousand loans can still be concentrated if all depend on one property market. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into portfolio limits. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

27. Granularity

Granularity is the degree to which no single exposure dominates the portfolio. Fine granularity supports law-of-large-numbers diversification of idiosyncratic risk. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Large N with small exposure shares is the idealised direction. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Granularity does not diversify systemic factor risk. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into IRB intuition. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

28. Credit migration

Credit migration is movement among rating or credit-quality states before default. It affects bond values and future PD even without immediate default. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Transition matrix P contains state-to-state probabilities. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Treating ratings as permanent until default ignores downgrade and upgrade risk. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into bond credit analysis. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

29. Transition matrix

Transition matrix is a matrix of probabilities of moving from each credit state to every other state over a horizon. It encodes migration and default together. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Each row sums to one. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Matrix powers assume a time-homogeneous Markov structure unless the model is altered. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into ratings. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

30. Absorbing default state

Absorbing default state is a default state that remains default within the transition model once entered. It simplifies migration mathematics. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. P(D→D)=1. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Recovery belongs to a separate economic process even if default is absorbing in the state model. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit migration. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

31. Credit score

Credit score is a model output used to rank or estimate credit risk from borrower information. It can be transformed into PD only if calibrated for a defined event and horizon. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Score→PD mapping may use logistic or monotonic calibration. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A ranking score is not automatically a probability. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into retail credit. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

32. Logistic PD model

Logistic PD model is a model mapping a linear predictor into a probability between zero and one. It is common for interpretable binary credit modelling. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PD=1/(1+e^(−z)), z=β_0+β’x. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Good discrimination does not guarantee good calibration. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into scorecards. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

33. Odds

Odds is the ratio p/(1−p) associated with probability p. Credit scorecards often operate linearly in log-odds. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. odds=p/(1−p). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Odds and probability are not interchangeable; 1:4 odds correspond to p=20%, not 25%. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into score scaling. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

34. Log-odds

Log-odds is the logarithm of the odds. Logistic regression makes log-odds linear in predictors. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. log(p/(1−p))=β_0+β’x. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Interpreting coefficients directly as probability changes is wrong because the logistic transform is nonlinear. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into PD models. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

35. Calibration

Calibration is agreement between predicted PDs and realised default frequencies over appropriate groups and horizons. It asks whether probabilities mean what they say. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Compare average predicted PD with observed default rate, allowing for sampling uncertainty and cycle effects. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A model can rank borrowers well yet be systematically over- or under-predictive. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into model validation. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

36. Discrimination

Discrimination is the ability to rank riskier obligors above safer ones. It is distinct from calibration. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. AUC/Gini/KS are common ranking metrics. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A model can have high AUC but badly biased PD levels. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit scoring. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

37. Brier score

Brier score is mean squared error of binary probability forecasts. It is a proper scoring rule that reflects calibration and resolution. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Brier=mean((p_i−y_i)^2). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Class imbalance and base rates affect interpretation. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into PD validation. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

38. ROC AUC

ROC AUC is area under the receiver-operating-characteristic curve. It measures ranking performance across thresholds. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. AUC=P(score_default>score_nondefault) under one interpretation. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. AUC says nothing directly about probability calibration. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into discrimination. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

39. Gini coefficient

Gini coefficient is a common rescaling of AUC in credit scoring. It offers a familiar ranking metric. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Gini=2×AUC−1 in the standard relation. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Different sample definitions can make Gini comparisons misleading. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into model monitoring. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

40. KS statistic

KS statistic is maximum separation between cumulative score distributions for defaulted and non-defaulted cases. It is another rank-separation measure. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. KS=max|F_bad(s)−F_good(s)|. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Threshold selection and sample mix affect the statistic. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

41. Point-in-time PD

Point-in-time PD is a PD intended to reflect current borrower and macroeconomic conditions. It tends to move with the cycle. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PD_t=f(current state, borrower data). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using PIT PD where a through-the-cycle regulatory concept is required can misalign the framework. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into IFRS 9. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

42. Through-the-cycle PD

Through-the-cycle PD is a PD designed to smooth or average cyclical variation to a greater degree. It supports more stable rating philosophy in some systems. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Long-run calibration is often central. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. TTC does not mean economically invariant; definitions vary by institution and framework. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

43. Rating philosophy

Rating philosophy is the design choice governing how quickly ratings respond to cyclical or borrower changes. It shapes migration and PD dynamics. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PIT and TTC are endpoints of a spectrum, not binary universal categories. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Comparing migration rates across different philosophies without adjustment is misleading. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit governance. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

44. Macroeconomic overlay

Macroeconomic overlay is an adjustment linking credit parameters to economic scenarios or forecasts. It introduces forward-looking conditions into expected-loss estimates. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PD/LGD/EAD may be scenario-specific functions of unemployment, rates, GDP or property prices. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Overlaying expert judgment without governance can become an unexplained plug. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into IFRS 9 ECL. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

45. Scenario weighting

Scenario weighting is probability weights assigned to alternative forward-looking economic paths. It produces a probability-weighted expected loss across scenarios. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. ECL=Σ_s w_s ECL_s, Σw_s=1. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using only a central scenario can miss nonlinear loss response. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

46. 12-month ECL

12-month ECL is the portion of lifetime expected credit losses associated with default events possible within the next 12 months under IFRS 9 terminology. It is not simply cash shortfalls expected during the next 12 months. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Measure full cash shortfall consequences of defaults occurring in the 12-month default window. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Treating Stage 1 ECL as only 12 months of lost payments misstates the concept. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into accounting impairment. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

47. Lifetime ECL

Lifetime ECL is expected credit losses from all possible default events over the expected life of the instrument under IFRS 9. It expands the default window while still probability-weighting cash shortfalls. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Integrate or sum marginal default probabilities × conditional shortfalls × discount factors across life. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Summing unconditional annual PDs without survival adjustment can overstate lifetime probability. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into Stage 2 and Stage 3. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

48. Significant increase in credit risk

Significant increase in credit risk is the IFRS 9 trigger concept for moving from 12-month to lifetime ECL in the general model. It compares credit risk at reporting date with initial recognition subject to the standard’s requirements and practical expedients. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Assess changes in lifetime default risk and relevant qualitative information. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A mechanical days-past-due threshold alone may be insufficient as a complete SICR policy. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into IFRS 9 staging. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

49. Stage 1

Stage 1 is performing exposures under the IFRS 9 general model without significant increase in credit risk since initial recognition, subject to the framework’s details. It carries 12-month ECL. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Interest recognition is generally on gross carrying amount under the standard’s general model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Stage numbers are accounting categories, not borrower moral labels. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into ECL. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

50. Stage 2

Stage 2 is exposures with significant increase in credit risk but not necessarily credit-impaired. It generally carries lifetime ECL under the IFRS 9 general model. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Lifetime marginal PDs become relevant to measurement. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Stage 2 is not equivalent to default. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

51. Stage 3

Stage 3 is credit-impaired exposures under IFRS 9. It generally carries lifetime ECL and changes the basis of interest recognition. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Cash-flow shortfall estimates become central. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Stage 3 accounting treatment should not be casually equated with Basel default without checking definitions. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into impaired assets. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

52. Cash shortfall

Cash shortfall is the difference between contractual cash flows due and cash flows expected to be received, including timing effects. It is the direct IFRS 9 economic object underlying ECL. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. CreditLoss=PV(contractual cash flows−expected cash flows). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. PD×LGD×EAD is an implementation architecture, not the only permitted conceptual representation of ECL. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

53. Time value in ECL

Time value in ECL is the requirement to discount expected credit losses rather than simply add nominal shortfalls. Delayed payment can create credit loss even when full nominal amount is ultimately recovered. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PV(shortfall)=Σshortfall_t/(1+EIR)^t under a simplified representation. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Ignoring timing can understate loss from delayed recoveries. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into accounting. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

54. Effective interest rate discounting

Effective interest rate discounting is the discounting anchor used in IFRS 9 ECL measurement subject to the standard’s detailed rules. It keeps impairment linked to original effective economics. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Discount expected shortfalls using the required EIR basis. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using a current market credit spread as the ECL discount rate without basis in the standard changes the measurement concept. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

55. Basel IRB

Basel IRB is the regulatory framework in which approved banks use internal or supervisory credit-risk parameters within prescribed risk-weight functions. It separates model inputs from regulatory capital formulas. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PD, LGD, EAD and sometimes M enter CRE31/CRE32 formulas. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A bank’s internal expected-loss pricing model is not automatically its regulatory IRB model. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

56. Foundation IRB

Foundation IRB is an IRB approach where specified parameters are supervisory rather than fully bank-estimated. It creates a different ownership of inputs than advanced approaches. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Banks estimate PD while supervisory treatment governs other components for eligible exposures, subject to current rules. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Calling all IRB models ‘internal models for every parameter’ is inaccurate. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into Basel. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

57. Advanced IRB

Advanced IRB is an IRB approach allowing approved own estimates for additional risk components subject to restrictions and floors. It requires strong data, governance and validation. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Own-estimate LGD/EAD use is subject to framework eligibility and supervisory approval. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Advanced does not mean unconstrained; floors and restrictions apply. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

58. Regulatory PD floor

Regulatory PD floor is a lower bound on PD used in regulatory formulas for specified exposure classes. It prevents extremely low model PDs from driving capital toward zero. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Apply current Basel floor according to exposure class and in-force version. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using a future-version floor in a current-period calculation without checking effective date is wrong. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

59. Regulatory LGD floor

Regulatory LGD floor is a lower bound on LGD for specified exposures in IRB calculations. It constrains optimistic severity estimates. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Apply collateral- and exposure-specific floors from the applicable rule set. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Economic workout LGD and regulatory LGD may differ because purposes differ. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

60. Maturity parameter

Maturity parameter is effective maturity M used in specified regulatory credit-risk functions. It recognises that longer commitments can create more risk. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. M enters the prescribed risk-weight formula where applicable. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Final contractual maturity and Basel effective maturity are not always identical. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into IRB. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

61. Risk-weighted assets

Risk-weighted assets is a regulatory exposure measure produced by applying prescribed credit-risk weights or formulas. It connects credit risk to minimum capital ratios. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. RWA depends on framework, exposure class and approach. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. RWA is not expected loss and not accounting carrying value. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into bank capital. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

62. Provision

Provision is an accounting allowance for expected credit losses under the applicable accounting framework. It absorbs expected impairment through earnings/carrying value mechanics. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Provision/ECL measurement depends on accounting rules, not solely Basel capital. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Equating provisions with regulatory capital creates category errors. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into bank accounting. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

63. Capital

Capital is financial resources intended to absorb unexpected and other losses under prudential requirements. It is not interchangeable with provisions. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Capital ratio=eligible capital/RWA under prudential definitions. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Expected loss can be priced/provisioned while capital addresses a different loss layer. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into prudential regulation. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

64. Credit spread

Credit spread is yield compensation relative to a benchmark reflecting expected loss, risk premia, liquidity and other effects. It embeds market credit information before default. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Spread can be linked to hazard and recovery in reduced-form models. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Spread/PD approximations such as spread≈PD×LGD work only under restrictive small-probability simplifications. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into bond credit. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

65. Recovery-of-par model

Recovery-of-par model is a defaultable-bond assumption in which recovery is a fraction of par or claim amount. It is one possible recovery convention. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Recovery=R×par upon default under model rules. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Different recovery conventions produce different bond values and hazard calibrations. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit pricing. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

66. Recovery-of-market-value model

Recovery-of-market-value model is a convention where recovery is a fraction of pre-default market value. It changes the recursion in credit pricing. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Post-default value=R×pre-default value in the simplified model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. It is a modelling assumption, not a universal legal recovery process. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into reduced-form models. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

67. Structural credit model

Structural credit model is a model linking default to firm asset value relative to liabilities. It treats equity and debt within a balance-sheet option framework. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Merton-style default at horizon occurs when asset value falls below debt threshold. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Real capital structures and early default mechanisms are more complex. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into corporate credit. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

68. Reduced-form model

Reduced-form model is a model specifying default arrival through hazard/intensity processes rather than firm-value barrier mechanics. It is flexible for market calibration. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Survival S(t)=exp(−∫λ(u)du) under a simple intensity setup. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Calibrated intensities can reflect risk premia as well as physical default expectations. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

69. Physical PD

Physical PD is a probability intended to describe real-world default frequency under a statistical measure. It supports forecasting, provisioning and risk depending on design. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Estimate from observed/default-model data with stated horizon and conditioning. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Market prices do not directly reveal physical PD without risk-premium assumptions. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into risk management. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

70. Risk-neutral default probability

Risk-neutral default probability is a pricing probability consistent with market prices under a risk-neutral measure and assumed recovery/model. It supports valuation rather than real-world frequency prediction. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Calibrate hazard to credit spreads under pricing assumptions. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Calling risk-neutral PD a forecast of actual defaults confuses measures. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into derivatives and bonds. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

71. CVA

CVA is a valuation adjustment for counterparty credit risk on derivatives and other exposures. It uses time-varying exposure and default probabilities rather than a static EAD only. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. CVA≈ΣDF_t×EE_t×LGD×marginalPD_t in a simplified unilateral discretisation. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Ignoring wrong-way risk can understate CVA when exposure rises as credit worsens. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into counterparty credit. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

72. Expected exposure

Expected exposure is the expected positive future replacement value of a derivative or netting set. It is a dynamic exposure analogue to EAD in counterparty risk. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. EE(t)=E[max(V_t,0)] after relevant netting/collateral assumptions. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Average exposure alone misses tail PFE and wrong-way dependence. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into CVA. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

73. Wrong-way risk

Wrong-way risk is dependence where exposure is high when counterparty default risk is high. It breaks simple independence between EAD/exposure and PD. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. E[Exposure×Default] exceeds E[Exposure]E[Default] under adverse positive dependence. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Multiplying average exposure by unconditional PD can understate joint loss. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

74. Collateral

Collateral is assets or claims pledged to reduce loss severity or exposure. It can reduce LGD or EAD depending on framework and mechanics. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Net exposure=max(Exposure−recognised collateral,0) in a toy model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Collateral value can fall in stress and liquidation can take time. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit risk mitigation. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

75. Haircut

Haircut is a reduction applied to collateral value to reflect volatility, liquidity and liquidation risk. It prevents full nominal collateral credit. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. RecognisedCollateral=MarketValue×(1−haircut) in a simple model. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A haircut is not expected loss; it is a risk-control valuation adjustment. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into secured lending. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

76. Guarantee

Guarantee is a third-party promise that can shift or reduce credit risk subject to enforceability and framework rules. It changes the source of repayment support. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Loss depends on borrower default and guarantor performance. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Treating an unenforceable or highly correlated guarantor as risk-free creates false mitigation. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

77. Seniority

Seniority is the contractual ranking of claims in default or restructuring. It materially affects recovery expectations. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Higher priority can support lower LGD, all else equal. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Seniority alone does not determine recovery because collateral, enterprise value and jurisdiction matter. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

78. Covenant

Covenant is a contractual condition designed to constrain borrower behaviour or trigger protections. It changes default and recovery pathways. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Covenants can alter transition probabilities and restructuring leverage. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Counting covenants without analysing enforceability and headroom is superficial. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into corporate lending. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

79. Watchlist

Watchlist is a governance category for exposures requiring heightened monitoring. It can precede default and support earlier intervention. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Triggers may combine quantitative and qualitative signals. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Watchlist status is not a probability by itself. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit operations. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

80. Arrears

Arrears is past-due contractual payments. It is a strong but lagging credit signal. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Days-past-due can enter staging, collection and default processes. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Not all defaults begin with arrears; some failures are sudden. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

81. Vintage analysis

Vintage analysis is tracking cohorts of loans originated in the same period. It reveals underwriting and macro effects by origination cohort. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Compare cumulative default/loss curves by months-on-book. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Mixing vintages can hide deterioration in newer production. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into portfolio monitoring. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

82. Roll-rate analysis

Roll-rate analysis is measuring transitions among delinquency states over time. It supports collections and loss forecasting. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Estimate P(state_j next period | state_i today). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Roll rates change across economic regimes and collection policies. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

83. Cure rate

Cure rate is probability or frequency that delinquent/defaulted accounts return to a performing state under the governed definition. It affects lifetime loss and collections economics. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Cure=P(return to performing | delinquent/default state). Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A temporary payment can be mistaken for sustainable cure. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into collections. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

84. Write-off

Write-off is accounting or operational removal of an amount deemed uncollectible under policy. It is not necessarily the same moment as economic default. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Timing depends on accounting, tax and policy rules. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Using write-off date as default date can distort PD timing. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into loss data. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

85. Recovery lag

Recovery lag is time between default and recovery cash flows. It affects discounted LGD materially. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Longer lag lowers PV of otherwise identical nominal recovery. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Ignoring recovery lag understates loss from slow legal processes. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into LGD. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

86. Collateral volatility

Collateral volatility is uncertainty in collateral value between measurement and liquidation. It creates stressed recovery risk. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Stressed collateral value can be modelled with haircuts or scenarios. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Current appraisal value is not guaranteed liquidation value. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into secured credit. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

87. Foreclosure cost

Foreclosure cost is legal, administrative and transaction costs of enforcing and selling collateral. It reduces net recovery. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Net proceeds=SalePrice−Costs. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Ignoring costs is especially material for property and small-ticket secured loans. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into mortgage LGD. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

88. Prepayment

Prepayment is early repayment before contractual maturity. It changes exposure profiles and lifetime ECL cash flows. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. EAD and survival projections should reflect prepayment where relevant. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Ignoring prepayment can overstate lifetime exposure. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into loans. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

89. Refinancing risk

Refinancing risk is risk that borrower cannot replace maturing debt. It increases default probability near funding cliffs. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. PD can depend on debt-maturity schedule and market access. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A low leverage ratio does not remove short-term liquidity/default risk. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

90. Debt-service coverage

Debt-service coverage is cash-flow capacity relative to debt-service obligations. It is a borrower-level credit signal. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. DSCR=available debt-service cash flow/debt service under defined numerator. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Definitions vary and can be manipulated by optimistic forecasts. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into underwriting. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

91. Loan-to-value

Loan-to-value is exposure relative to collateral/property value. It links leverage to recovery and default incentives. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. LTV=loan balance/property value under chosen valuation basis. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. LTV is dynamic because both balance and property value change. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into mortgage credit. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

92. Probability-weighted scenarios

Probability-weighted scenarios is multiple economic paths with explicit weights. They capture nonlinear ECL response better than a single central forecast. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. ECL=Σw_sECL_s. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Weights and scenarios require governance because small changes can materially move provisions. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

93. Model overlay

Model overlay is a governed adjustment applied when models do not adequately capture known risks. It acknowledges model limitations explicitly. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Final estimate=ModelOutput+Overlay under policy. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Permanent unexplained overlays can become substitutes for fixing weak models. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into ECL governance. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

94. Backtesting PD

Backtesting PD is comparing predicted probabilities with realised defaults. It tests calibration and stability. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Observed default rate versus average PD, with confidence intervals and segmentation. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. One good aggregate year can hide severe subgroup miscalibration. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into validation. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

95. Backtesting LGD

Backtesting LGD is comparing predicted loss severity with realised discounted workout losses. It tests recovery models. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Compare estimated LGD at reference date with final/updated workout LGD. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Open default cases create censoring and incomplete recovery histories. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

96. Backtesting EAD

Backtesting EAD is comparing predicted default exposure with realised balances at default. It tests utilisation and CCF assumptions. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Realised CCF=(EAD−drawn reference balance)/undrawn commitment under a defined denominator. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Facility cancellations and limit changes complicate denominators. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

97. Low-default portfolio

Low-default portfolio is a portfolio with too few defaults for straightforward statistical estimation. It creates parameter uncertainty and validation challenges. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Use pooling, external data, Bayesian methods or conservative frameworks as appropriate. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. No defaults observed does not imply zero PD. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into wholesale credit. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

98. Margin of conservatism

Margin of conservatism is an explicit conservative adjustment for uncertainty or data/model deficiencies under governed frameworks. It prevents false precision. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Parameter_final=estimate+MoC under the applicable methodology. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Conservatism should be traceable to uncertainty rather than arbitrary padding. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into model governance. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

99. Data lineage

Data lineage is the traceable path from source data through transformations into model inputs and outputs. It is essential because credit parameters are only as reliable as the data definitions beneath them. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Store source, timestamp, transformation and version metadata. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A calibrated formula on inconsistent default flags is still wrong. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into model risk. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

100. Model drift

Model drift is deterioration in model behaviour as borrower populations, policies or economic relationships change. It makes periodic validation necessary. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Monitor calibration, discrimination, input distributions and residuals through time. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Stable code does not imply stable model performance. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit monitoring. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

101. Population drift

Population drift is change in the characteristics of borrowers or exposures entering the model. It can invalidate development-sample relationships. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Compare feature/score distributions and segment mix over time. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. Improved or worsened applicant quality can be mistaken for model change. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

102. Policy drift

Policy drift is change in underwriting, collections, limit management or recovery practice. It alters observed outcomes even if borrower risk is unchanged. Credit models become reliable only when the event, horizon, population and economic meaning are specified with the number.

Mathematics. Version operational policy alongside model performance data. Units matter: PD and LGD are proportions, EAD is currency, maturity is time, and discounted recovery depends on both cash amount and date. A model that multiplies percentages correctly but mismatches horizons is still wrong.

Failure mode. A PD model calibrated before a major policy change may become misaligned. Jo’s diagnostic is to identify which component owns the failure: event probability, conditional severity, exposure size, discounting, dependence or framework definition. That prevents a vague “credit model error” from hiding the repair.

Connection. This concept feeds into credit systems. Ryan would verify it against realised outcomes or an independent cash-flow calculation; Aisha would ask whether stress conditions change the same input in the same direction as other credit variables. The aim is a loss system whose assumptions can be audited.

Worked Example 1: The Basic PD × LGD × EAD Identity

A S$500,000 term loan has a one-year PD of 1.5%, LGD of 35% and projected one-year EAD of S$420,000. The simple one-year expected loss is 0.015×0.35×420,000=S$2,205.

The result is an average loss under the model, not the loss that will literally appear on this borrower. In a two-state simplification the borrower either does not default, producing no default loss, or defaults and produces a loss around S$147,000 under deterministic LGD/EAD. The expected value averages these states.

Verification: if PD doubles while LGD and EAD remain fixed, EL doubles. If collateral reduces LGD from 35% to 20%, EL falls proportionally. These comparative statics should be visible before calculating.

Worked Example 2: Recovery Timing Changes LGD

At default, EAD is S$1,000,000. Case A recovers S$650,000 net immediately. LGD is 35%. Case B recovers the same nominal S$650,000 three years later. At a 5% annual discount rate, recovery PV is 650,000/(1.05)^3≈S$561,500, so economic LGD is roughly 43.85% before any additional workout cost.

The nominal recovery rate looks identical in both cases. The economic loss does not. This is why serious LGD models preserve recovery cash-flow timing rather than only total recovery dollars.

Ben’s verification is to set the discount rate to zero. The delayed case should then collapse back toward the undiscounted 35% LGD.

Worked Example 3: Revolving EAD

A credit line has S$40,000 drawn and S$60,000 undrawn at the reference date. A teaching EAD model assumes 50% of undrawn commitment will be drawn before default. EAD=40,000+0.50×60,000=S$70,000.

If PD=4% and LGD=60%, expected loss is 0.04×0.60×70,000=S$1,680. Using current drawn balance only would give S$960 and understate loss by 43% in this simplified example.

The example shows why EAD is a behavioural forecast for revolving products, not simply a balance lookup.

Worked Example 4: Multi-Period Lifetime Expected Loss

Suppose a three-year loan has marginal default probabilities 1.0%, 1.5% and 2.0% in years 1, 2 and 3, with EADs S$100,000, S$80,000 and S$55,000, LGD 40%, and annual ECL discount factor based on 4%. A simplified lifetime EL is the sum of each year’s marginal default probability × LGD × EAD × discount factor.

Year 1 contribution ≈0.010×0.40×100,000/1.04. Year 2 contribution ≈0.015×0.40×80,000/1.04². Year 3 contribution ≈0.020×0.40×55,000/1.04³. Sum the three contributions.

The important point is not the arithmetic. The PDs used here are marginal default probabilities associated with default in each period. Plugging cumulative PDs into every year would double count defaults.

Worked Example 5: Conditional PD and Survival

Assume conditional annual PDs are 1%, 2% and 4% for the next three years. Survival through year 1 is 0.99. Survival through year 2 is 0.99×0.98=0.9702. Survival through year 3 is 0.9702×0.96=0.931392.

Three-year cumulative PD is 1−0.931392=6.8608%. It is not 1%+2%+4%=7% because the later conditional PDs apply only to survivors.

The marginal year-3 default probability is survival through year 2 times year-3 conditional PD: 0.9702×0.04=3.8808%. This is the quantity that belongs in a period-by-period expected-loss sum.

Worked Example 6: Bayes and a Credit Warning Signal

Suppose 2% of comparable borrowers default within a year. A warning signal is present in 70% of eventual defaulters and 10% of non-defaulters. In 10,000 borrowers, expect 200 defaults: 140 signal-positive. Among 9,800 non-defaulters, 980 are signal-positive. Total positive signals=1,120.

Posterior default probability given the signal is 140/1,120=12.5%. The signal multiplies risk substantially but does not imply a 70% PD. The difference is the base-rate effect.

This is why credit monitoring needs calibrated posteriors, not merely classifier sensitivity.

Worked Example 7: Expected Loss Does Not Measure Tail Concentration

Portfolio A has 1,000 loans of S$10,000 each, each with 1% PD and 50% LGD. Portfolio B has one S$10 million loan with 1% PD and 50% LGD. Both have expected loss S$50,000 under the simple model.

Their loss distributions are radically different. Portfolio B either loses roughly S$5 million or nothing under deterministic LGD. Portfolio A can diversify idiosyncratic defaults if borrower events are not strongly correlated.

Expected loss alone therefore cannot measure concentration or unexpected loss.

Worked Example 8: Correlation Changes Tail Loss, Not Expected Loss

Take 100 identical loans with PD 2%, LGD 50% and EAD S$100,000. Expected loss is 100×0.02×0.50×100,000=S$100,000 regardless of whether defaults are independent or correlated, provided marginal PDs remain 2%.

But high positive dependence makes multi-default scenarios much more likely. The 99.9th percentile portfolio loss can increase dramatically even while EL remains exactly S$100,000.

This is the mathematical boundary between average loss and capital/tail risk.

Worked Example 9: Loan-to-Value and LGD Stress

A property loan has EAD S$800,000 and property collateral valued at S$1,000,000, giving current LTV 80%. Suppose stress reduces property value by 25% to S$750,000 and liquidation costs are 8% of sale proceeds. Net collateral proceeds before timing discount are S$690,000.

Ignoring other recoveries, nominal loss becomes S$110,000, or LGD 13.75%. If realisation takes two years and is discounted, economic LGD rises further. The example demonstrates why current LTV is not itself LGD.

Collateral value, liquidation discount, cost and time all sit between LTV and realised loss.

Worked Example 10: IFRS 9 Scenario Weighting

A bank models three economic scenarios for a portfolio: base ECL S$8m with 60% weight, upside ECL S$5m with 15% weight, downside ECL S$18m with 25% weight. Probability-weighted ECL=0.60×8+0.15×5+0.25×18=S$10.05m.

The weighted ECL is higher than base-case ECL because the downside scenario is nonlinear and severe. Replacing the three scenarios with the base case alone would understate the probability-weighted result by S$2.05m in this toy example.

The example also shows why scenario design and weights are model inputs requiring governance, not afterthoughts.

Worked Example 11: PD Calibration by Grade

Suppose a rating grade has 5,000 borrower-years with average predicted one-year PD 1.2% and 72 observed defaults, an observed rate of 1.44%. The raw difference is 24 basis points.

A validation team should not instantly declare failure or success. It should consider sampling uncertainty, rating mix, cycle conditions, default-definition consistency and whether observations are independent. Larger samples allow tighter conclusions than low-default portfolios.

Calibration is statistical evidence, not visual closeness alone.

Worked Example 12: Credit Spread Approximation

In a very simple one-period small-probability intuition, a credit spread may be approximated by PD×LGD if risk premia, liquidity, discounting and timing are ignored. With PD 2% and LGD 60%, the expected-loss component is about 1.2% of exposure over the horizon.

A market spread of 250bp should not therefore be interpreted as a 4.17% physical PD by dividing 2.5% by 60% without further assumptions. The spread can include risk premia, liquidity, market technicals and differences between risk-neutral and physical probabilities.

This is an important boundary between pricing mathematics and forecasting mathematics.

Expected Loss, Pricing, Provisioning and Capital Are Four Different Jobs

A lender can use expected credit loss as one input to loan pricing: the interest margin needs to cover funding, operating cost, expected credit loss, capital cost and target return, among other items. Accounting provisioning uses the applicable accounting standard. Regulatory capital follows prudential rules. Economic capital may use an internal loss-distribution framework. The numbers can share PD/LGD/EAD concepts without being interchangeable.

This separation is essential. If an internal pricing PD is point-in-time and an IRB PD is long-run calibrated, the two can legitimately differ. If IFRS 9 LGD includes forward-looking recovery assumptions and accounting discounting, it may differ from regulatory downturn LGD. The correct response is reconciliation, not forced equality.

Clara’s governance rule is that every parameter should carry a purpose label as clearly as it carries a percentage.

IFRS 9: Cash Shortfalls First

IFRS 9’s core ECL concept is a probability-weighted present value of cash shortfalls: contractual cash flows due minus cash flows expected to be received, considering timing and reasonable/supportable information. PD×LGD×EAD is a common modelling architecture because it decomposes the same economic problem, but implementations must still satisfy the accounting measurement objective.

This matters for delayed payments. If a borrower eventually pays the full nominal amount but substantially later than contractually due, the time-value shortfall can still be a credit loss. A pure undiscounted default/no-default model can miss that effect.

The mathematics therefore returns to time value of money: credit risk is not only how much cash arrives, but when.

Basel IRB: Expected Loss and Unexpected Loss Are Separated

The Basel IRB architecture uses PD, LGD, EAD and, for specified exposures, effective maturity within prescribed risk-weight functions. The framework explicitly separates expected loss treatment from the capital requirement designed for unexpected loss. This is conceptually important even for readers who never implement the regulatory formula.

Expected loss is the centre of a loss distribution. Prudential capital is designed around adverse loss beyond ordinary expectation under regulatory assumptions. A bank that priced only expected loss but held no loss-absorbing capital would be vulnerable to perfectly foreseeable statistical variation around the mean.

That distinction is one of the cleanest bridges between probability theory and bank capital mathematics.

PD Is a Forecast Target, Not a Personality Score

A probability of default is about an event over a horizon, not a moral judgement about a borrower. Good modelling defines the population, observation date, outcome window, default event and allowed predictors before fitting an equation. Changing any of those changes the meaning of the output.

Two lenders can assign different PDs to the same borrower because they use different horizons, products, information, calibration windows or default definitions. The correct question is not which number “is the true PD” in the abstract, but which model is appropriate and calibrated for the decision.

This is why explainability begins with target definition before feature importance.

LGD Is a Recovery Cash-Flow Model

LGD is often taught as 1−recovery rate. That shortcut hides most of the real work. Recovery can arrive from borrower payments, collateral sales, guarantors, insurance, restructurings and legal proceedings. Costs can include legal fees, servicing, property maintenance, taxes and collection expense. Timing can span months or years.

A robust LGD process therefore reconstructs workout cash flows, discounts them under the governed methodology, segments by collateral/seniority/product, addresses incomplete recovery cases, and tests downturn behaviour. The output is a percentage, but the model underneath is a dated cash-flow system.

The existing specialist article How Banks Estimate Loss Given Default owns the deeper implementation mechanics.

EAD Is a Behaviour Model

For a fully drawn bullet loan, EAD can look easy. For a credit card, overdraft, revolving line or contingent commitment, it is behavioural. Borrowers under stress may draw unused limits before default. Banks may reduce limits, facilities may expire, and interest/fees may accumulate. The path matters.

That is why EAD models often use credit conversion factors, utilisation trajectories, segmentation and reference-date definitions. A model should distinguish contractual limit, available undrawn amount, current drawn balance and realised exposure at default.

A single balance column cannot answer all four questions.

Credit Risk Is Conditional on the Economy

Default and recovery are cyclical. Recession can reduce borrower cash flow while simultaneously depressing collateral values. That makes PD and LGD positively dependent in bad states: more borrowers default and each default can be more severe. EAD can also rise when borrowers draw liquidity before failure.

A simple model that stresses PD alone while freezing LGD and EAD can therefore understate downturn loss. Scenario-based credit mathematics needs coherent co-movement: employment, rates, property prices, corporate earnings and refinancing conditions should affect the relevant components consistently.

This is one reason forward-looking ECL and bank stress testing are systems problems rather than isolated parameter shocks.

A Professional Credit-Risk Workflow

Define the product, obligor, facility and legal exposure.
Define default precisely and freeze the historical mapping rules.
Choose the PD horizon and rating/score philosophy.
Estimate or assign PD with calibration and uncertainty controls.
Model EAD from contractual amortisation, utilisation and contingent commitments.
Model LGD from collateral, seniority, recovery timing and workout costs.
Construct expected loss at the correct horizon and discount basis.
Build migration, scenario and concentration views where needed.
Separate pricing, accounting, regulatory and economic-capital uses.
Backtest PD, LGD and EAD independently before relying on combined EL.
Stress common macro factors and dependence among borrowers.
Version data, models, overlays and approvals so every number is reproducible.

Common Failure Modes

1. Multiplying mismatched horizons

A one-year PD combined with lifetime EAD/LGD without a transition model creates an incoherent EL. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

2. Using default rate as borrower PD

Observed cohort frequency is data; borrower PD is a forward model output conditional on scope and information. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

3. Using 1−recovery as LGD mechanically

Recovery timing and workout costs can make economic LGD higher. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

4. Using current balance as EAD for revolving credit

Borrowers can draw more before default. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

5. Adding annual PDs

Later default probabilities apply only to survivors; use survival/marginal mathematics. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

6. Treating EL as capital

Expected loss and unexpected-loss capital solve different problems. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

7. Treating provision as capital

Accounting allowances and prudential capital have different frameworks. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

8. Using market spread as physical PD

Market spreads include premia, liquidity and pricing-measure effects. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

9. Freezing LGD in stress

Recoveries often deteriorate when defaults rise. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

10. Ignoring concentration

Same EL can hide radically different tail losses. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

11. Ignoring calibration

A beautifully ranked score can produce wrong probability levels. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

12. No purpose label

Regulatory, IFRS 9, pricing and economic parameters can legitimately differ. The repair is to restate the event, horizon, conditional structure and model purpose before recomputing the loss.

Formula Map

Concept	Simplified formula	Meaning
One-period loss	L=EAD×LGD×I(default)	Realised default loss in a simple deterministic-severity model.
Expected loss	EL=PD×LGD×EAD	Average one-period loss under deterministic LGD/EAD.
Recovery-linked LGD	LGD=(EAD−PV(net recoveries))/EAD	Economic severity after discounted net recovery.
EAD with CCF	EAD=Drawn+CCF×Undrawn	Simplified contingent-exposure projection.
Survival	S_T=Π(1−q_t)	Probability of surviving conditional PDs through T.
Cumulative PD	1−S_T	Default probability by horizon T.
Marginal PD	S_{t−1}q_t	Unconditional probability of default during period t.
Portfolio EL	ΣPD_iLGD_iEAD_i	Linear sum of individual expected losses.

Authoritative Reference Map

Connected Bukit Timah Tutor Credit-Risk Route

Applied Case Study 1: A Singapore housing loan

Situation. A mortgage has declining balance, property collateral, borrower income risk and potentially long recovery time after default. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Project EAD from amortisation, estimate PD by horizon, model stressed property recovery and workout costs, then discount recovery cash flows. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Current LTV is an input, not the final LGD. Property-value and unemployment stress can move PD and LGD together. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 2: A revolving credit card

Situation. A cardholder has a S$20,000 limit but only S$6,000 currently drawn. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Estimate PD from a calibrated account model, model pre-default utilisation and CCF for EAD, and use unsecured retail LGD assumptions appropriate to the purpose. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Current balance alone can understate exposure if distressed customers draw remaining limits. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 3: An SME working-capital line

Situation. The borrower depends on receivables and seasonal cash flow and can draw before default. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Model utilisation by months-on-book and risk grade, estimate collateral/guarantee recovery and stress refinancing conditions. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Facility behaviour and borrower PD should not be estimated independently if stress draws are risk-related. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 4: A corporate term loan

Situation. The borrower has scheduled amortisation and a refinancing cliff at maturity. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Build time-varying EAD from contractual payments, term PD/hazard from financial conditions, and LGD from enterprise/collateral recovery. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Low near-term PD can coexist with high maturity-refinancing risk. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 5: A bond portfolio

Situation. Credit deterioration affects market value before default. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Separate migration/spread risk from terminal default loss and use transition matrices or credit-spread scenarios in addition to EL. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. PD×LGD×EAD alone does not describe mark-to-market spread loss. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 6: A low-default sovereign or bank portfolio

Situation. Historical defaults are sparse. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Use external data, long histories, expert structure, conservative uncertainty treatment and validation methods appropriate to low-default portfolios. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Zero observed defaults is evidence of rarity, not evidence of zero probability. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 7: A secured commercial-property loan

Situation. Collateral covers the loan at origination but property prices fall in recession. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Stress collateral value, sale discount, workout cost and liquidation time jointly with borrower PD. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Collateral can become least protective precisely when defaults are most frequent. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 8: A credit-card EAD backtest

Situation. Modelled CCF is 40%, but realised utilisation before default is rising. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Compare realised EAD and CCF by vintage, delinquency state and score band; recalibrate or investigate policy drift. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. EAD drift can come from borrower behaviour or lender limit-management changes. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 9: An IFRS 9 Stage 1 portfolio

Situation. Loans have not experienced significant credit-risk increase. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Calculate 12-month ECL using default events possible in the next 12 months but measure the full cash shortfalls resulting from those defaults. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Stage 1 ECL is not simply twelve months of missed instalments. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 10: An IFRS 9 Stage 2 portfolio

Situation. Credit risk has increased significantly since initial recognition. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Use lifetime default probabilities, scenario-weighted LGD/EAD/cash shortfalls and the required discounting basis. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Stage 2 is not default; lifetime ECL reflects the expanded default window. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 11: A bank stress test

Situation. A recession raises unemployment, corporate failures and property losses. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Apply coherent macro paths to PD, LGD and EAD, aggregate portfolio losses and compare with provisions/capital resources. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Independent single-parameter shocks can understate nonlinear joint stress. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

Applied Case Study 12: A concentrated borrower group

Situation. Several large obligors share one sector and parent-supplier network. The purpose is to identify which component—PD, LGD, EAD, dependence or horizon—actually owns the uncertainty.

Method. Preserve single-name exposures, common-factor dependence and concentration rather than relying only on average PD. Adrian maps exposure dates, Jo defines default and horizon, Aisha checks recovery economics, and Ryan recomputes expected loss under at least one alternative scenario.

Boundary. Same total EL can hide much higher tail loss when exposures are concentrated. Mira then asks whether the parameter is being used for pricing, accounting, regulatory capital or internal risk. The same symbol can carry different governed definitions across those uses.

A 20-Question Credit-Risk Diagnostic

What exact event counts as default?
What is the PD horizon?
Is PD physical, regulatory, accounting, pricing or risk-neutral?
Is PD marginal, conditional or cumulative?
What population and product does the model cover?
How is EAD expected to change before default?
Are undrawn commitments included?
What recovery cash flows are included in LGD?
Are workout costs deducted?
Are recoveries discounted?
Does LGD reflect downturn conditions where required?
How does collateral value behave in stress?
Do PD and LGD worsen together under common macro factors?
Is the portfolio concentrated by obligor, sector or geography?
Are model parameters calibrated and backtested separately?
Are incomplete defaults handled correctly in LGD data?
Are scenario weights and overlays governed?
Can the result be reconciled to cash-flow shortfalls where relevant?
Which framework owns the number: pricing, IFRS 9, Basel, economic capital or management risk?
What evidence would trigger recalibration or model replacement?

Final Principle

Credit risk is not one probability. It is a loss architecture: default likelihood × conditional severity × exposure, embedded in time, dependence, recovery and governance.

The simple PD×LGD×EAD identity is valuable because it exposes the anatomy of expected loss. The world-class version of the model then asks whether each component uses the right horizon, whether survival is handled correctly, whether recoveries are discounted, whether exposure grows before default, whether common stress moves parameters together, and whether the number belongs to accounting, pricing, capital or another decision.

That progression is what turns credit risk from a spreadsheet multiplication into financial mathematics. It also creates clean boundaries for the next lanes: bank balance sheets, bank capital, bank liquidity, deposits and loan pricing, stress testing, market risk and systemic risk.

Small Group Tutorials

Banking And Finance Mathematics | Credit Risk, PD, LGD, EAD and Expected Loss Mathematics