Jump-diffusion (Merton 1976 / Zhou 1997)¶
Merton’s 1976 option-pricing extension adds a compound-Poisson jump to the asset process; Zhou (1997) repurposed it for structural credit modelling. The resulting default probability has a clean Poisson-weighted series form that subsumes the standard Merton model.
Setup¶
Under the risk-neutral measure:
with
Na Poisson process of intensity\lambda(mean jumps per year),Y \sim \text{lognormal}(\mu_J, \sigma_J^2)the jump multiplier,\kappa = e^{\mu_J + \sigma_J^2/2} - 1 = E[Y] - 1chosen so the total drift matchesr - q.
Conditional Gaussian¶
Conditional on exactly n jumps in [0, T], the asset log-return
is Gaussian:
with
PD formula¶
Unconditioning on N_T:
where
The implementation truncates the series at \lambda T + 6\sqrt{\lambda T}
jumps, which captures \geq 1 - 10^{-12} of the Poisson mass.
Example¶
from merton import Firm
from merton.extensions import JumpDiffusionModel, jump_diffusion_pd
# Direct PD: Merton baseline + 0.5 expected jumps/year, downward bias
pd = jump_diffusion_pd(
asset_value=100, asset_vol=0.30, debt=60, rf=0.04, T=1.0,
jump_intensity=0.5, jump_mean=-0.05, jump_std=0.15,
)
# Model usage
firm = Firm(equity=50, debt_short=30, debt_long=50, equity_vol=0.5, horizon=1.0)
result = JumpDiffusionModel(
jump_intensity=1.0, jump_mean=-0.10, jump_std=0.20,
).fit(firm)
print(result.summary())
When to prefer jump-diffusion¶
Crisis-period calibration. Crash risk that Merton can’t capture (LTCM 1998, 2008) shows up as a fat negative tail in equity returns. Adding negative-mean jumps fits observed credit spreads at short horizons much better.
Emerging-market sovereign analytics. Currency-crisis “jumps” are the dominant source of default risk; modelling them explicitly matters.
Long-horizon LGD assumptions. Jumps push the expected loss higher even when the average drift is unchanged.
The trade-off is parameter estimation: \lambda, \mu_J, and
\sigma_J require equity-options data or judgement-based priors. For
firms without listed options, fixing them to industry medians is a
reasonable default.