Forecasting Default with the Merton Distance to Default Model
Source: Bharath, S. T. & Shumway, T. (2008) · Review of Financial Studies 21(3), 1339–1369 · doi:10.1093/rfs/hhn044
TL;DR
Evaluates how well the structural Merton distance-to-default (DD) model predicts corporate default. The headline finding: Merton DD is not a sufficient statistic for default. A simple "naive" approximation — which keeps the Merton functional form but skips the iterative nonlinear solve — performs slightly better in hazard models and out-of-sample than both the full Merton DD model and a reduced-form model using the same inputs. An expanded hazard model with additional predictors beats Merton DD probabilities out of sample. The value lies in the functional form (leverage and volatility), not in solving the structural equations.
Problem it solves
Merton (1974) treats equity as a call option on firm assets, so default occurs when asset value falls below the face value of debt; "distance to default" is how many asset-volatility units the firm sits above that boundary, and the implied default probability is N(−DD). The question: is this elegant structural solve actually the best way to forecast default, or is it the form that matters?
The method
Merton DD (π_Merton): infer unobserved asset value V and asset volatility σ_V from equity value, equity volatility, and debt by an iterative procedure (Crosbie–Bohn / Vassalou–Xing): seed σ_V = σ_E·E/(E+F), back out daily asset values and implied log returns, re-estimate σ_V and drift μ, iterate until |Δσ_V| < 10⁻³; then π = N(−DD), DD from Equation (7).
Naive DD (no equation solving): set naive D = F (face value of debt); naive σ_D = 0.05 + 0.25·σ_E; naive σ_V = (E/(E+F))·σ_E + (F/(E+F))·(0.05+0.25·σ_E); set expected asset return naive μ = r_{i,t−1} (firm's own prior-year equity return). Then naive DD = [ln((E+F)/F) + (r_{i,t−1} − 0.5·naive σ_V²)·T] / (naive σ_V·√T), with π_naive = N(−naive DD).
Comparison: Cox proportional hazard models and out-of-sample forecasts; also tests against KMV-style alternatives and CDS-/bond-implied default probabilities.
Assumptions & inputs
Inputs: equity value, equity volatility, face value of debt (Compustat), CRSP daily returns. Sample: NYSE/AMEX/NASDAQ non-financial firms, 1980–2003, with default data from Altman's database (1980–2000, supplemented 2001–2003); 1,016,552 firm-months, variables winsorized at the 1st/99th percentiles.
Maintained Merton assumptions: lognormal asset value, single debt horizon, debt as a zero-coupon bond at face value F.
How to use it
For ranking firms by default risk, the naive measure is a cheap, robust substitute that captures essentially the same signal (leverage + volatility scaled by horizon).
In practice, combine DD (or naive DD) with market and accounting variables in a hazard model rather than relying on the structural probability alone.
Limitations & pitfalls
Default is rare and clustered; samples are imbalanced and regime-dependent.
Structural assumptions (lognormal assets, single horizon, F as strike) are stylized; the authors note the naive volatility/return choices are deliberately ad hoc, made to isolate the form from the solve.
The full KMV implementation (empirical DD distribution, proprietary debt adjustments) is not exactly replicable here.
Key references
Bharath, S. & Shumway, T. (2008) — Forecasting Default with the Merton Distance to Default Model — Review of Financial Studies
Merton, R. (1974) — On the Pricing of Corporate Debt: The Risk Structure of Interest Rates — Journal of Finance
Shumway, T. (2001) — Forecasting Bankruptcy More Accurately: A Simple Hazard Model — Journal of Business
Provenance: verified/generated from the paper's full text.
