TL;DR
This paper empirically tests five leading structural models for pricing corporate bonds, finding that while they capture some aspects of bond pricing, none perfectly predict market spreads. The study highlights that existing models struggle with both under- and over-predicting spreads, particularly for safer or riskier bonds, suggesting a need for models that better balance these extremes.
What Problem Does This Paper Address?
Corporate bonds, unlike government bonds, carry the risk of default. Structural models attempt to price this risk by linking bond prices to the underlying financial health and capital structure of the issuing firm. However, a persistent challenge has been that these models often fail to generate the wide range of credit spreads observed in the market. Some models predict spreads that are too low, while others overstate them, especially for firms with high leverage or volatility. This paper aims to empirically evaluate five prominent structural models to understand their performance, identify their systematic biases, and pinpoint areas for improvement in corporate bond pricing.
Methods
The authors implement and compare five structural models: the Merton (1974) model, the Geske (1977) model, the Leland and Toft (1996) model, the Longstaff and Schwartz (1995) model, and the Collin-Dufresne and Goldstein (2001) model. They use a "portfolio of zeroes" approach for coupon bonds in the Merton, Longstaff-Schwartz, and Collin-Dufresne-Goldstein models, while the Geske model uses a compound option approach for coupons. The study estimates model parameters using historical financial data and market data. They then analyze prediction errors in terms of pricing errors, yield errors, and, most importantly, credit spread errors. To understand systematic biases, they conduct t-tests and multivariate regression analyses of prediction errors against various firm and bond characteristics.
Datasets
The study uses a sample of 182 noncallable corporate bonds from firms with simple capital structures between 1986 and 1997. Bond prices are obtained from the Fixed Income Database. Financial data (balance sheets, accounting data, stock prices, dividend yields, share repurchases) are sourced from COMPUSTAT and CRSP. Interest rate data are from the Constant Maturity Treasury (CMT) series reported by the Federal Reserve Board.
Key Insights and Ideas
- No Perfect Model: All tested structural models exhibit substantial prediction errors, indicating that none fully capture the complexity of corporate bond pricing.
- Spread Extremes: A common issue across models is the tendency to generate extremely low spreads for "safe" bonds (low leverage, low volatility) and extremely high spreads for "risky" bonds.
- Geske Model Improvement: The Geske model, with its endogenous default boundary and compound option approach to coupons, shows improvement over the simpler "portfolio of zeroes" approach used in other models.
- LT Model Overprediction: The Leland and Toft model consistently overpredicts spreads, largely due to its continuous coupon assumption, a surprising finding given the literature.
- Stochastic Interest Rates Matter (but with caveats): The Longstaff-Schwartz and Collin-Dufresne-Goldstein models, which incorporate stochastic interest rates, generally produce higher predicted spreads but are sensitive to interest rate volatility estimates.
- Leverage is Key: Leverage is a consistently significant predictor of spread prediction errors across most models, highlighting its central role in credit risk.
- Maturity's Limited Role: Contrary to some literature, maturity does not appear to be a primary driver of systematic prediction errors once other factors like bond rating are controlled for.
Key Results
- Average Errors Misleading: The average spread prediction error is a poor indicator of model performance due to large dispersion.
- Merton & Geske: Tend to underpredict spreads on average, with Geske showing less severe underprediction.
- Leland & Toft: Overpredicts spreads on average, by more than twice the observed market spread for the average bond.
- Longstaff-Schwartz & Collin-Dufresne-Goldstein: Generate higher predicted spreads than Merton and Geske but at the cost of significant accuracy issues and extreme dispersion.
- Systematic Errors: T-tests reveal systematic errors related to leverage, bond rating, and maturity (especially for the Merton model). Regression analysis confirms leverage and asset volatility as major determinants of prediction errors.
Limitations and Future Work
- Data Limitations: The study focuses on firms with simple capital structures, which may not be representative of all corporate bond issuers. The exclusion of callable bonds also limits the scope.
- Parameter Sensitivity: The accuracy of structural models is sensitive to parameter estimates, particularly for asset volatility and leverage.
- Model Simplifications: Models make simplifying assumptions about recovery rates, default boundaries, and interest rate dynamics that may not fully reflect reality.
- Future Research: The authors suggest future research should focus on developing models that can raise spreads on safer bonds without excessively increasing them for riskier bonds. Better modeling of coupons and recovery rates, and more accurate interest rate models (beyond Vasicek) are also recommended.
Summary updated