To derive bankruptcy predictors, the authors use the least absolute shrinkage and selection operator (LASSO). They show that in addition to common market value predictors, accountancy data can predict future default risk and that the importance of accountancy-based variables increases as forecasting horizons get longer. Furthermore, LASSO performs well in out-of-sample datasets, which is something that classic bankruptcy models struggle with.
Investing in corporate bonds implies default risk; hence, investors need to be aware of early warning signs. Merton (Journal of Finance 1974) developed a distance-to-default model, and since then, several studies have been performed based on either market or accountancy indicators. The predictive power of these models, however, tends to be weak, so the authors assess their model with both in- and out-of-sample tests. They compare their results with the previous studies by Shumway (Journal of Business 2001) and Campbell, Hilscher, and Szilagyi (CHS, Journal of Finance 2008). Of Shumway’s original five predictive variables, least absolute shrinkage and selection operator (LASSO) uses four variables (either identical or reshaped). Five of the eight CHS variables also enter the LASSO equation. Furthermore, LASSO selects two additional predictive variables. The authors show that the LASSO model presents consistent predictive variables across different time periods.
How Is This Research Useful to Practitioners?
Variable selection methods are used to identify predictive variables and to improve the predictability of future bankruptcy. Previous studies have used such accountancy variables as Altman’s Z-score and Ohlson’s Z-score. Shumway’s market variable model and CHS’s research changed the focus on market-based variables.
In a departure from the approach of previous researchers, the authors apply LASSO, in which predictive variables are selected based on a shrinkage algorithm. They create a unique US dataset consisting of both accounting-based and market-based data. The model is then tested with both in- and out-of-sample datasets. Several goodness-of-fit measures are used (such as Akaika, R2, and Brier score), the results of which indicate that the LASSO-selected variables are statistically significant. The authors then conclude that LASSO performs better than the CHS model in both the in- and out-of-sample datasets. Furthermore, they show that the distance-to-default variable (a key variable in Merton’s model) has less predictive power than the LASSO variables.
Lastly, the authors determine whether the variables selected by LASSO are consistent for long- and short-term forecast horizons. When the forecasting period increases, the accountancy variables become more significant.
Because many financial professionals are involved in either creating excess return or avoiding unnecessary risks, a good understanding of predictors for future default is important. This research provides useful insight for a wide range of financial professionals.
How Did the Authors Conduct This Research?
LASSO uses a shrinkage method, which provides a sparse variable set solution. It is a method that is easy to interpret and has the advantage of model selection stability and prediction accuracy. It is also computationally efficient and avoids the multicollinearity problem, which is particularly handy in terms of bankruptcy parameters because these market and accountancy variables are strongly correlated. In total, 39 variables are considered to be predictors for future bankruptcy.
The authors combine annual Compustat accounting data and daily and monthly CRSP equity data from 1980 to 2009 to create a US default dataset. To align the market and accountancy data, the accountancy information is lagged by four months. Default is defined as a company filing for Chapter 7 or Chapter 11 bankruptcy protection. Defaulted firms are assigned a bankruptcy indicator of 1, and firms that have not endured bankruptcy or are removed from the dataset (due to mergers and acquisitions) are assigned a 0. The dataset shows that bankruptcy filings are strongly countercyclical with peak periods.
A logistic regression model is used for a one-year-ahead default prediction model. The stepwise subset selection method is applied to derive the most predictive variables. A finite sample size goodness-of-fit method is used to assess model fitting. The authors demonstrate that leverage has predictive powers in two ways: Market leverage shows the firm’s ability to pay, and book leverage shows the precautionary actions taken by management. Because market leverage is more sensitive to stock price changes, book leverage is a less vulnerable measure and has stronger predictive powers.
Default risk affects a wide range of investment decisions. Both asset managers and advisers create asset mixes consisting of several levels of risk and return. For potential additional return from risky assets to be assessed, the additional default risk needs to be measured and the potential for future default needs to be monitored. Most historical default models have the disadvantage that only the difference between asset and liability values is taken into account and little predictive power is derived. The authors create a unique US market and accountancy dataset and test their model’s predictive power by using different goodness-of-fit measures and different datasets. Because many financial professionals’ work is affected by default risk, this research will be interesting for a wide range of people working in finance.