Merton Credit Model: Equity, Debt and Default Risk

Payoffs and risk-neutral values

Robert Merton's 1974 paper On the Pricing of Corporate Debt: The Risk Structure of Interest Rates starts with a deliberately spare balance sheet. A firm has assets and one zero-coupon promise. If the assets cannot cover that promise at maturity, creditors take the assets.

Suppose the assets are worth £120 million, the firm promises £100 million in one year, annual asset volatility is 25% and the continuously compounded risk-free rate is 4%. The model values equity at about £26.66 million and risky debt at £93.34 million. Equity is worth more than the £20 million accounting residual because shareholders keep the upside and have limited liability.

That is Merton's structural insight: equity is a European call option on firm assets. It links equity volatility, leverage, default probability and credit spreads in one model. It is also easy to misuse, especially when a risk-neutral pricing probability is presented as a real-world forecast.

At time $t$, let:

• $V$ be the market value of firm assets;
• $F$ be the face amount of zero-coupon debt due at $T$;
• $\tau=T-t$ be years to maturity;
• $\sigma_V$ be annualised asset volatility;
• $r$ be a constant continuously compounded annual risk-free rate.

All values must use the same currency. Rates and volatility are annualised decimals, and $\tau$ is measured in years.

The basic model assumes frictionless trading, no taxes, no bankruptcy costs, no payouts from the asset pool, constant $r$ and $\sigma_V$, and no default before $T$. Under the physical measure $P$,

$$\frac{dV_s}{V_s}=\mu\,ds+\sigma_V\,dW_s^P.$$

At maturity, equity and debt pay

$$E_T=(V_T-F)^+, \qquad B_T=\min(V_T,F)=F-(F-V_T)^+.$$

Equity is a call with strike $F$. Risky debt is a risk-free zero-coupon bond minus a put on firm assets. Under the risk-neutral measure $Q$, and with no asset payout,

$$\frac{dV_s}{V_s}=r\,ds+\sigma_V\,dW_s^Q.$$

The Black-Scholes formula then gives

$$\begin{aligned} d_1^Q&= \frac{\ln(V/F)+(r+\tfrac12\sigma_V^2)\tau} {\sigma_V\sqrt{\tau}},\\ d_2^Q&=d_1^Q-\sigma_V\sqrt{\tau},\\ E&=V\Phi(d_1^Q)-Fe^{-r\tau}\Phi(d_2^Q). \end{aligned}$$

Because the simple firm contains only debt and equity, $V=E+B$. Therefore

$$B=V\Phi(-d_1^Q)+Fe^{-r\tau}\Phi(d_2^Q) =Fe^{-r\tau}-P_{BS}.$$

The first term represents asset value delivered in default states and the second the promised payment in survival states. Recovery is endogenous: creditors receive $V_T$, not a fixed recovery fraction.

The risk-neutral terminal default probability is

$$Q(V_T<F)=\Phi(-d_2^Q).$$

This is the probability consistent with prices inside the model. It is not automatically a physical probability suitable for credit approval.

A worked one-year balance sheet

For $V=120$, $F=100$, $\tau=1$, $r=0.04$ and $\sigma_V=0.25$,

$$d_1^Q=1.0143, \qquad d_2^Q=0.7643.$$

Substitution gives $E\approx26.66$ and $B\approx93.34$, in the same monetary units as $V$ and $F$. The model-implied continuously compounded debt yield is

$$y=-\frac{1}{\tau}\ln\left(\frac{B}{F}\right) \approx6.89\%,$$

so the model credit spread is about $6.89\%-4.00\%=2.89\%$, or 289 basis points. The risk-neutral probability $\Phi(-d_2^Q)$ is about 22.2%.

Those figures are internally consistent, not a valuation recommendation. A 22.2% pricing probability does not mean 22.2% of comparable firms will default. Risk premia and the model's recovery mechanism both enter debt value.

If the physical expected asset return is instead $\mu=8\%$, physical distance to the maturity threshold is

$$DD_P= \frac{\ln(V/F)+(\mu-\tfrac12\sigma_V^2)\tau} {\sigma_V\sqrt{\tau}} =0.9243,$$

and the model physical default probability is

$$P(V_T<F)=\Phi(-DD_P)\approx17.8\%.$$

$DD_P$ is not $d_2^Q$ unless the physical and risk-neutral drifts happen to match. Calling both quantities "$d_2$" conceals the most important measure distinction in the model.

Inferring unobserved assets

Public equity value and equity volatility are observable, but $V$ and $\sigma_V$ are not. Itô's lemma gives the equity volatility relation

$$\sigma_E E =\frac{\partial E}{\partial V}\sigma_VV =\Phi(d_1^Q)\sigma_VV.$$

Together with the call-price equation, this supplies two equations for $V$ and $\sigma_V$ from observed $E$, $\sigma_E$, $F$, $r$ and $\tau$. In the worked example, model equity volatility is about 95.1%. Leverage makes equity much more volatile than the 25% asset process.

This implementation solves both equations in log parameters, which keeps asset value and volatility positive:

import numpy as np
from scipy.optimize import root
from scipy.stats import norm

def equity_outputs(V, sigma_V, face, years, rate):
    scale = sigma_V * np.sqrt(years)
    d1 = (np.log(V / face) + (rate + 0.5 * sigma_V**2) * years) / scale
    d2 = d1 - scale
    equity = V * norm.cdf(d1) - face * np.exp(-rate * years) * norm.cdf(d2)
    return equity, d1, d2

def infer_assets(equity, sigma_E, face, years, rate):
    def residual(log_params):
        V, sigma_V = np.exp(log_params)
        model_E, d1, _ = equity_outputs(V, sigma_V, face, years, rate)
        model_sigma_E = norm.cdf(d1) * sigma_V * V / equity
        return [model_E / equity - 1.0, model_sigma_E / sigma_E - 1.0]

    guess = np.log([equity + face * np.exp(-rate * years), 0.20])
    result = root(residual, guess)
    if not result.success:
        raise RuntimeError(result.message)
    return np.exp(result.x)  # V, sigma_V

The face amount is only a proxy when a real company has loans, leases, coupons and many maturities. Sensitivity analysis around $F$ and $\tau$ is more honest than pretending the balance sheet supplies an exact option strike.

Another implementation reconstructs a daily asset-value series for a trial $\sigma_V$, estimates volatility from its log returns and iterates. That can reduce dependence on a single noisy equity-volatility input, but it is a different estimator and needs convergence and data-quality checks.

Merton, distance to default and KMV

Merton's one-bond model and the commercial KMV methodology are related, not identical. The public KMV account by Peter Crosbie and Jeffrey Bohn, Modeling Default Risk, describes three broad steps:

1. infer asset value and asset volatility from equity and liability data;
2. measure distance from asset value to a chosen default point;
3. map distance to an empirical Expected Default Frequency (EDF) using historical defaults.

The commonly quoted KMV default-point approximation is short-term debt plus half of long-term debt. It is not Merton's single face value $F$, and it should not be presented as the full proprietary production methodology. The default point, asset drift treatment and empirical EDF map have all been implementation choices rather than consequences of the 1974 formula.

The empirical map is important. Applying $\Phi(-DD)$ assumes the physical asset process is lognormal with a known drift. An EDF map instead asks how often firms at similar measured distances actually defaulted. That can absorb regularities missed by the normal model, though it also introduces calibration-sample, definition and regime dependence.

Distance to default is therefore a score before it is a probability. Its horizon, debt proxy, drift convention and calibration vintage should travel with the number.

Extensions and model risk

Merton permits default only at the debt maturity. Fischer Black and John Cox's 1976 first-passage model adds a covenant barrier that can trigger default earlier. Hayne Leland's 1994 model and the Leland-Toft 1996 extension connect endogenous bankruptcy to taxes, bankruptcy costs and debt maturity.

Continuous diffusion also makes default over a very short horizon negligible for a solvent firm whose threshold remains below current assets. Observed short-maturity spreads need not vanish because they contain jump-to-default risk, liquidity, taxes, risk premia and market frictions. Chunsheng Zhou's 2001 jump-diffusion credit model addresses the first of those mechanisms. Our Merton jump-diffusion guide derives the related asset process.

The basic model has further limits:

• one zero-coupon liability cannot represent a real maturity ladder or covenant package;
• firm asset value is latent, so model outputs depend on an inversion;
• constant volatility and lognormal assets exclude stochastic volatility and jumps;
• traded equity may be illiquid or absent, making the main signal weak;
• debt spreads reflect liquidity and risk premia as well as expected default loss;
• early restructuring and strategic default do not fit a single terminal test.

Reduced-form models take a different route: they specify a default intensity and recovery process without deriving default from firm assets. They often calibrate more directly to bond or CDS curves, while structural models give a clearer link to leverage and equity. "Better" depends on whether the task is economic interpretation, physical default prediction or market pricing.

For multi-name credit, dependence adds another layer. A copula model can join firm-level default distributions, but it cannot repair a poor default threshold or an unstable asset estimate.

Merton's model lasts because its balance-sheet logic is sharp. Its output is only as literal as its one-bond firm.

Frequently Asked Questions

Is Merton distance to default equal to Black-Scholes d2?

Only after specifying a measure. Pricing $d_2^Q$ uses the risk-neutral drift $r$; physical distance to default uses a physical asset drift or another forecasting convention. They are generally different numbers.

What debt number should be used as the strike?

The original model has one promised face amount at one maturity. A practical proxy may combine current and long-term liabilities, but there is no exact strike for a real maturity ladder. Document the proxy and test sensitivity to it.

Why is equity volatility higher than asset volatility?

Equity is a levered residual claim. In the model, $\sigma_E=\Phi(d_1^Q)\sigma_VV/E$, so a small equity value relative to assets magnifies volatility.

Does Merton produce a real-world default probability?

It can under a physical asset process and a specified drift, giving $\Phi(-DD_P)$. Debt prices instead use risk-neutral dynamics, while KMV-style systems map distance to empirically observed default frequencies.