Pages: 26

Distress risk and corporate failure modelling research has developed into a prodigious literature spanning more than six decades. Financial crises, such as the Asian Financial Crises of 1997, the “tech wreck” of 2001, the global financial crisis (2007–2009) (GFC), and the COVID pandemic (from early 2020) have led to an unprecedented number of corporate collapses that, to different degrees, have seriously impacted the global economy. Financial crises and their contagion effects underscore the complex interdependencies of the global financial system. While corporate failure used to be considered the domain of small and/or newly listed companies, the failure of large, well-established corporations is now commonplace. The importance of accurate and robust corporate failure prediction models has never been more apparent. This book examines the strengths and limitations of several traditional corporate failure modelling techniques, such as LDA, logit/probit, neural networks, CART, and hazard models. However, the focus of this book is on the potential of modern machine learning methods, such as gradient boosting machines, adaptive boosting (AdaBoost), random forests, and deep learning to advance the corporate failure prediction field.

Pages: 49

The distress risk and corporate failure prediction literature is very extensive. However, it can be divided into two broad streams of research. The first are statistical learning approaches, which use some type of statistical model to predict failure based on observed data and a set of features or explanatory variables. The second stream involves theoretically derived models such as gambler’s ruin models and the KMV distance to default measures. Over the decades, a wide range of corporate failure modelling techniques has emerged, including simple linear models such as LDA, logit, and probit models. However, the failure prediction literature has moved on to more sophisticated discrete choice models, such as mixed logit, nested logit, and hazard models. There is a long-standing debate in the literature about the predictive value of different explanatory variables, particularly accounting vs market price variables. Some studies have tried to show that market price variables are highly predictive (particularly more sophisticated measures such as the KMV distance-to-default measure); however, the consensus view seems to be that using a combination of accounting and market-based variables produces more optimal predictive models. There is good evidence that including asset volatility measures into failure models can improve explanatory and predicative power of distress and failure forecasting models. Among the many financial ratios used in failure forecasting, measures based on cash flows, leverage, and working capital appear to have the strongest predictive power overall. However, there are many other features that may have predictive power, such as macro-economic variables and earnings management and corporate governance proxies.

Pages: 34

This chapter introduces a range of machine learning methods, including neural networks, CART, MARS, gradient boosting, AdaBoost, and random forests. Of all the tree-based machine learning methods, the gradient boosting model seems to consistently provide the strongest out-of-sample predictive performance across many studies. The model also has good interpretability through relative variable importance score, partial dependency plots, and interaction effects (as will be shown in Chapter 4). Deep learning is another potentially powerful method that can be applied to distress risk and corporate failure research. While the literature is still very limited, a recent study based on US data shows that a deep learning model outperforms a hazard model on Type I classification performance. Modern machine learning methods offer many advantages for corporate failure research, particularly given some of the general characteristics of corporate failure datasets. For instance, machine learning methods such as gradient boosting are more able to handle “dirty data” issues, such as non-normalness, missing values, outliers, database errors, and so on. Machine learning methods are also designed to handle situations where the number of predictors is greater than the sample size (p>n), which is useful considering that corporate failure samples tend to be quite small, while the number of potential predictors of failure can be very large. Because modern machine learning methods require minimal researcher intervention, they can reduce “snooping bias” and p-hacking, which has been gaining much attention in recent literature. And since modern machine learning methods are high dimensional, they can be used to test the predictive power of many hundreds, if not thousands, of features without compromising model stability and performance.

Pages: 40

This chapter provides a comprehensive illustration of the various statistical learning methods described in Chapter 2 and Chapter 3. In particular, Chapter 4 focuses on the estimation and interpretation of classification and regression trees (CART), multivariate adaptive regression splines (MARS), generalized lasso, gradient boosting machines, and random forests. The software used for this illustration is Salford Predictive Modeler (Version 8.2), which has impressive diagnostic and data visualization capability, particularly compared to open-source software such as R. The gradient boosting model was the best performance model based on the AUC. The model was 94.03% accurate on Type I errors (predicting a failed firm to be safe) and 93.82% accurate on Type II errors (predicting a safe firm to fail). The gradient boosting, random forests, and CART models all revealed slightly different variable importances. However, 23 of the 26 features were found to be important (i.e., had an RVI > 0) in all models. Financial, market price, and corporate governance proxies (such as shareholder ownership and concentration) contributed to the overall predictive power of the models. Model diagnostics are very important in machine learning models as the model architecture settings can influence empirical results. For instance, the adjustable parameters of a gradient boosting model include learn rates, number of trees, maximum nodes per tree, minimum number of terminal nodes, choice of loss function, and so on. Chapter 4 demonstrates several diagnostic and robustness tests, including model stability tests.

Pages: 41

Much of the distress risk and corporate failure research has focused on public companies. There is comparatively little distress risk research devoted to private companies, not-for-profits, and public sector entities. With respect to private company distress research, there is a wide diversity of failure definitions, sample sizes, modelling techniques, use of explanatory variables, reporting jurisdictions, and data sources. Similar to public companies, private company distress studies have used mainly conventional discrete choice models, such as LDA and logit/probit models. However, there has been recent research that has used machine learning methods in binary and multi-class settings, such as Jones and Wang (2019). As with private companies, there are very good reasons to develop distress predictions for not-for-profits. One of the key reasons is that the not-for-profit sector is socially and economically significant to the entire global economy. However, as with private companies, there is quite limited research on the distress of not-for-profit entities. Public sector distress models face the same challenges. The continual development and refinement of distress prediction models for private sector entities, not-for-profits, and public sector entities remains a major challenge and opportunity for future research.

Pages: 7

While this volume has reviewed a wide range traditional distress risk and corporate failure modelling techniques such as LDA, logit/probit, neural networks, and hazard models, I have drawn particular attention to modern machine learning methods that hold much promise in this area of research. For instance, machine learning models have significantly better predictive power than conventional models, on both Type I and Type II error rates. These models can also be interpreted through relative variable importance (RVI) scores, partial dependency plots, and interaction effects. Machine learning methods are also better equipped to handle common data issues, such as non-normalness, missing values, database errors, and so on. Machine learning models can reduce biases associated with data snooping and p-hacking. Three potentially fruitful areas of future research are: (1) the inclusion of a wider range of predictor variables to model distress risk and corporate failure, such as macro-economic variables, earnings management variables, and corporate governance proxies; (2) to utilize the latest advances in text mining and natural language processing (NLP). Future research can utilize innovative techniques in natural language processing (NLP) to exploit the predictive value of signal embedded in unstructured text data; and (3) more research is needed to develop more accurate and robust distress prediction models for private companies, not-for-profit entities, and public sector entities.