Reader question: Black–Scholes assumes continuous price paths. What changes mathematically if an asset can jump suddenly before the next hedge can be placed?
Merton’s jump-diffusion model keeps the continuous Brownian-motion part of Black–Scholes and adds a compound-Poisson jump process. The result is a return distribution with a continuous core plus occasional discontinuous moves. For European options under the classic lognormal-jump specification, the price can be written as a Poisson-weighted mixture of Black–Scholes-type prices.
This article owns the discontinuous one-factor option-pricing problem: diffusion volatility + jump intensity + jump-size distribution + option terms → a risk-neutral jump-diffusion distribution, European option values, smile effects, simulation outputs and diagnostics about whether the jump specification earns its extra complexity.
It does not own stochastic-volatility calibration, generic Lévy models, local-volatility surfaces or structural credit modelling. Those are separate mathematical roles. The purpose here is to show exactly what a Poisson jump changes, how the drift is corrected, why the pricing series works and where jump risk breaks continuous hedging intuition.
This is public mathematical and computational education. It is not personalized financial advice or a recommendation to trade options.
1. The diffusion-only baseline
Under Black–Scholes with continuous dividend yield q, the risk-neutral asset dynamics are:
dSt/St = (r−q)dt + σdWt.
The sample path is continuous. Over an infinitesimal interval, price changes can be arbitrarily small. Dynamic delta hedging relies on that continuity.
Real markets can move discontinuously after announcements, policy shocks, defaults or overnight information. Jump diffusion asks what happens when the continuous-path assumption is deliberately relaxed.
2. Add a compound-Poisson jump process
A common risk-neutral specification is:
dSt/St− = (r−q−λκ)dt + σdWt + (J−1)dNt.
Here:
- Wt is Brownian motion;
- Nt is a Poisson process with intensity λ;
- J is the multiplicative jump factor;
- κ = E[J−1] is the mean proportional jump;
- σ is diffusion volatility;
- r−q−λκ is the compensated risk-neutral drift.
The left-limit notation St− reminds us that a jump changes the price instantaneously from its pre-jump level.
3. Why the drift needs the term −λκ
If jumps have positive or negative mean size, they contribute to expected asset growth. Under risk-neutral pricing, the discounted asset price must have the appropriate martingale property.
Over a short interval dt, the expected proportional jump contribution is approximately:
λκ dt.
Subtracting λκ from the continuous drift offsets that expected jump contribution so the total risk-neutral expected growth remains r−q.
Omitting the compensator double-counts expected jump growth and violates the intended risk-neutral drift.
4. Lognormal jump sizes
In Merton’s classic specification, the log jump size is normal:
Y = ln J ~ N(μJ, δ²).
Therefore:
J = eY
and:
κ = E[J−1] = exp(μJ + ½δ²) − 1.
The three jump parameters have distinct roles:
- λ: how often jumps arrive;
- μJ: average log jump direction;
- δ: variation in jump sizes.
These parameters can trade off against one another during calibration, creating identifiability problems.
5. Conditional on the jump count, returns are Gaussian
Suppose exactly n jumps occur before maturity T. The sum of n normally distributed log jump sizes is also normal.
Therefore the terminal log return conditional on NT=n remains Gaussian, with mean and variance adjusted by n.
This is the key tractability result:
conditional on n jumps → Black–Scholes-like lognormal world.
The unconditional distribution is then a mixture over n=0,1,2,… weighted by Poisson probabilities.
6. Poisson probabilities determine the mixture weights
For a Poisson process with intensity λ:
P(NT=n) = e−λT(λT)n/n!.
If λT is small, the zero- and one-jump terms dominate. If λT is near or above one, multiple-jump terms can carry substantial probability mass.
A pricing implementation truncates the infinite series only after verifying that omitted Poisson mass is negligible for the desired accuracy.
7. The European-option mixture
Under the classic Merton specification, a European option can be expressed as an infinite weighted sum of Black–Scholes-type values:
V = Σn=0∞ wn BSn,
where wn is the relevant Poisson weight and BSn is a Black–Scholes value with parameters adjusted for the conditional n-jump distribution.
The exact parameter rearrangement depends on the chosen notation for jump size and risk-neutral jump intensity. A robust implementation derives the formula from one internally consistent convention rather than mixing formulas from different sources.
8. The no-jump limit is a mandatory benchmark
If:
λ → 0,
the probability of any jump disappears. The model must reduce to Black–Scholes with diffusion volatility σ.
This is an important software invariant.
Falsifier: set λ numerically close to zero and compare the jump-diffusion price with an independent Black–Scholes implementation. Material disagreement indicates a convention or coding error.
9. Jump intensity is not the same as jump probability
λ is the expected number of jumps per unit time. The probability of at least one jump by T is:
1 − e−λT.
For λT=1, that probability is about 63.2%, not 100%.
Confusing intensity with probability is a common interpretation error.
10. Diffusion variance and jump variance play different roles
σ controls continuous path variability. δ controls dispersion of jump sizes. Increasing either can raise option values, but the strike and maturity pattern differs.
Diffusion volatility affects every instant. Jump variance contributes through discrete jump events whose probability depends on λT.
Short-dated out-of-the-money options can therefore respond strongly to jumps because a single discontinuity can move the asset across the strike even when continuous diffusion has little time to act.
11. Why jumps create implied-volatility smiles and skews
A single Black–Scholes volatility produces one lognormal terminal distribution. A jump-diffusion model produces a mixture distribution with extra skewness and kurtosis.
When jump-diffusion option prices are inverted through the Black–Scholes formula, the resulting implied volatility varies with strike.
Negative average jumps tend to strengthen downside asymmetry; jump-size variance thickens tails. The exact implied-volatility shape depends on all jump and diffusion parameters together.
Matching a smile does not prove the jump mechanism is uniquely correct because stochastic-volatility and local-volatility models can generate similar price patterns.
12. Calibration is a nonlinear inverse problem
Suppose observed option prices or implied volatilities are yk and model values are mk(θ), with:
θ = (σ, λ, μJ, δ).
A calibration can minimise:
Σ wk[mk(θ) − yk]².
The weights may reflect price scale, bid–ask spreads, vega or another criterion.
Different objective functions can produce different parameters even when fitted prices look similar.
13. Parameter identifiability is a weak link
Higher λ with smaller jump sizes can resemble lower λ with larger jump sizes. Diffusion volatility can also substitute for some jump variance.
As a result, several parameter sets may generate similar option surfaces.
Falsifier: start the optimizer from many dispersed initial points and profile the objective around the optimum. If materially different parameter sets fit almost equally well, do not interpret the calibrated parameters as uniquely identified economic truths.
14. Short maturities often contain the strongest jump information
For short maturity, ordinary diffusion has less time to accumulate variance. A large one-off move therefore has a distinctive effect on far-from-the-money option prices.
But short-dated options can also be noisy, illiquid or strongly affected by event-specific information.
A calibration that depends entirely on a few short-dated quotes is fragile even if those quotes are theoretically informative about jumps.
15. Simulation algorithm
A direct terminal simulation can proceed as follows:
- draw N ~ Poisson(λT);
- draw a Brownian normal shock Z;
- draw N log-jump sizes Y1,…,YN;
- compute the compensated diffusion drift;
- sum the jump logs;
- construct ST;
- evaluate the payoff;
- discount and average across paths.
For path-dependent products, jump times must also be simulated or a time-stepping scheme must account for their occurrence within the path.
16. Exact jump-time simulation
Conditional on N jumps over [0,T], the unordered jump times are distributed like N independent uniforms over the interval; sorting them gives the event times.
Alternatively, interarrival times can be sampled as exponentials with rate λ.
These methods avoid forcing jumps to occur only on a fixed time grid.
17. PDE becomes PIDE
Under jump diffusion, option value does not satisfy only a local diffusion PDE. Jumps connect the value at S with values at remote states SJ.
The pricing equation contains an integral term over jump sizes, creating a partial integro-differential equation, or PIDE.
Numerical methods may split the local diffusion operator from the nonlocal jump integral, use FFT/convolution techniques or rely on transform methods when the characteristic function is tractable.
18. The characteristic function remains tractable
Jump diffusion belongs to the Lévy-process family with a known characteristic exponent. For log price, the characteristic function combines:
- the Gaussian diffusion exponent;
- the compound-Poisson jump exponent.
This makes Fourier methods attractive and connects directly to transform-based option-pricing algorithms.
19. Jump risk breaks perfect continuous hedging
In Black–Scholes, continuous trading in the underlying and cash can replicate a European option under the idealised assumptions.
When the asset jumps, the option value and stock price can move discontinuously before a delta hedge is rebalanced. A hedge chosen immediately before the jump generally cannot eliminate every possible jump-size outcome.
The market becomes incomplete unless additional assumptions or traded instruments span the jump risk.
This is not a numerical inconvenience. It is a structural change in the economics of replication.
20. What risk-neutral pricing assumes about jump risk
Because jump risk is not generally spanned by the stock alone, moving from a physical jump process to a risk-neutral jump process can require an assumption about the market price of jump risk.
Merton’s classic option-pricing setup makes a particular assumption that preserves tractability. Other models and calibrations can imply different risk-neutral jump parameters.
Therefore historical jump frequency need not equal the jump intensity calibrated from option prices.
21. Inputs and outputs
Inputs can include:
- spot price;
- strike and maturity;
- interest-rate and dividend assumptions;
- diffusion volatility σ;
- jump intensity λ;
- mean log jump μJ;
- jump volatility δ;
- risk-neutral convention;
- Poisson-series truncation rule;
- calibration instruments and weights;
- Monte Carlo seed if simulation is used.
Outputs can include:
- European option price;
- Black–Scholes benchmark;
- jump premium relative to the diffusion-only model;
- Poisson term contributions;
- implied-volatility smile;
- calibration residuals;
- parameter uncertainty;
- simulated jump counts and paths;
- hedge-error scenarios;
- series-convergence diagnostics.
22. Evidence polarity
Evidence for confidence includes exact recovery of Black–Scholes as λ approaches zero, stable Poisson-series convergence, small and pattern-free option-calibration residuals across strikes and maturities, reproducible parameters across nearby calibration windows, agreement between series, Fourier and Monte Carlo implementations, and improved out-of-sample pricing relative to a simpler diffusion benchmark.
Evidence against confidence includes unstable jump parameters, strong dependence on optimizer starting values, calibration concentrated in one maturity region, poor out-of-sample fit, series truncation sensitivity, contradictions between historical and risk-neutral interpretations that are not acknowledged, or hedge errors that reveal missing volatility or jump structure.
23. Counterexample: jumps fit the smile but are not uniquely required
A Merton model can fit a smile better than constant-volatility Black–Scholes. But a Heston stochastic-volatility model or local-volatility surface may fit the same quotes as well or better.
Falsifier: compare out-of-sample pricing and hedging across alternative model classes rather than declaring the jump mechanism proven by in-sample smile fit.
24. Counterexample: parameter degeneracy
Suppose one calibration returns λ=0.4 with large jump sizes and another returns λ=1.3 with smaller jumps, yet both price the calibration set almost identically.
Falsifier: inspect the objective-function surface and Hessian or profile likelihood. A flat valley indicates weak identification.
25. Counterexample: too few Poisson terms
For large λT, truncating the series after only a few jump counts can omit meaningful probability mass.
Falsifier: increase the truncation level until price changes fall below a stated tolerance and track the omitted Poisson tail probability.
26. Counterexample: wrong drift compensation
Mixing a jump-size convention for J with a compensator formula derived for log jump Y can silently bias prices.
Falsifier: independently verify the martingale condition numerically:
E[e−(r−q)TST] = S0
under the chosen risk-neutral convention, up to Monte Carlo error.
27. Counterexample: short-dated event risk dominates calibration
A scheduled earnings or policy event can create a strong one-maturity jump signature. Calibrating one constant jump process to that surface may export event-specific risk into all future dates.
Falsifier: remove the event maturity and recalibrate. If the inferred jump intensity or size changes radically, the constant-jump model is absorbing calendar-specific structure.
28. Counterexample: jump hedging is treated like diffusion hedging
A delta hedge that works well under small continuous moves can suffer a large instantaneous residual when the underlying jumps.
Falsifier: simulate discrete jump scenarios and calculate pre-jump hedge P&L. If the residual remains material even at arbitrarily high rebalancing frequency, it is jump incompleteness rather than ordinary discretisation error.
29. Diagnostic ladder
- No-jump limit: recover Black–Scholes as λ→0.
- Martingale test: verify drift compensation under the chosen convention.
- Poisson-mass test: track omitted series weight.
- Independent implementation: compare series, Fourier or Monte Carlo values.
- Multi-start calibration: detect local minima and parameter degeneracy.
- Residual map: inspect errors by strike and maturity.
- Parameter stability: compare rolling calibration windows.
- Model benchmark: compare Black–Scholes, Heston or local volatility.
- Out-of-sample test: price quotes excluded from calibration.
- Hedge-error test: stress discontinuous moves.
- Event-maturity test: isolate scheduled jump risk.
- Simulation-tail test: compare model return tails with relevant empirical evidence without conflating physical and risk-neutral measures.
30. Alternatives
Black–Scholes is simpler and appropriate as a baseline when continuous lognormal dynamics are adequate for the intended task.
Heston stochastic volatility creates smile and skew through random variance rather than price jumps.
Bates models combine stochastic volatility and jumps.
Kou double-exponential jumps use a different jump-size law with asymmetric exponential tails.
Variance Gamma and other Lévy models replace the compound-Poisson-plus-Brownian structure with different pure-jump or infinite-activity processes.
Local volatility fits a state-dependent diffusion surface without explicit jumps, although it has different forward-dynamics implications.
31. Connections to the surrounding Bukit Timah Tutor knowledge estate
The diffusion baseline is Black–Scholes option-pricing algorithms.
Observed option prices are inverted into volatility in implied-volatility solver algorithms.
A major alternative smile mechanism is Heston calibration.
Characteristic-function pricing connects to Carr–Madan FFT option-pricing algorithms, while general numerical PDE machinery is developed in finite-difference option-pricing algorithms.
Path simulation connects to Monte Carlo pricing algorithms.
The full lane is indexed at Finance & Banking Algorithms | Applied Mathematics in Real Financial Systems.
32. What would falsify confidence?
Confidence should be withdrawn if the model fails its no-jump Black–Scholes limit, drift compensation violates the martingale condition, Poisson truncation is material, parameters are not identifiable, calibration residuals show systematic strike/maturity structure, alternative models outperform out of sample, or the intended use assumes that delta hedging can eliminate jump risk when the model itself implies otherwise.
33. Verification and update triggers
Preserve the jump convention, risk-neutral drift formula, σ, λ, μJ, δ, calibration instruments, objective weights, optimizer settings, series truncation and independent benchmark prices for each model version.
Trigger review when the implied-volatility surface changes materially, event risk becomes important, calibration parameters become unstable, jump-count truncation expands, historical jump behaviour changes, the model is applied to a new asset class, or hedging errors reveal persistent structure not captured by the calibrated jump process.
34. Primary and high-quality references
- Robert C. Merton, Option Pricing When Underlying Stock Returns Are Discontinuous, Journal of Financial Economics, 1976.
- Robert C. Merton’s original working-paper version is archived through MIT Sloan / RePEc.
- For transform-based pricing connections, see the characteristic-function methods linked from the Bukit Timah Tutor Carr–Madan article and the original Fourier-option-pricing literature.
- Federal Reserve Board and OCC, SR 11-7 — Guidance on Model Risk Management, for general principles of conceptual soundness, validation, ongoing monitoring and model limitations.
Educational boundary: Jump diffusion is a model of discontinuous price dynamics, not proof that market jumps follow a Poisson clock or lognormal size law. Its value comes from making those assumptions explicit, computable and falsifiable.