Name

Prediction of Corporate Failure

Prepared for

Sorbonne Master of Financial Economics

This project was a partial replication of Financial Distress and Corporate Acquisitions: Further Empirical Evidence (Theodossiou, Panayiotis, et al. 1996). The authors created a sample of 181 healthy and distressed firms from 1981 to 1989, and collected financial ratios from COMPUSTAT. The authors implement two sequential logit models, one to predict risk of default, then among the positive class, another to predict corporate acquisition. I implement only the first model, and use the data as a sandbox to compare a number of traditional econometric and machine learning models.

The project has three particularly novel aspects. The first is the addition of iso-cost curves to ROC space. Using hypothetical costs of type 1 and type 2 errors, a point of tangency between a cost function and the ROC curve can be found, representing the best trade-off between the two. The exercise shows where economics has important contributions to make in the field of data science; most treatments of threshold selection in machine learning texts treat the issue abstractly, and give no guidance on how to select a proper threshold.

Second is the use of Kendall’s Tau to perform model selection. Using perturbation analysis, feature importance computed for all models. These are ordinal rankings, so Kendall’s Tau can be used to find the single model that captures the consensus of the whole ensemble.

Finally is a novel computation of AUC using concordant and discordant pairs. The AUC score can be shown to be precisely equal to the percentage of concordant pairs in the sample. This is shown in the project.

You can check it out here.