Early bankruptcy prediction of firms

Abstract

Bankruptcy prediction is as important a problem for Banks, Governments, Investors & Entrepreneurs as it is challenging from an Analytics point of view. Historically, it has been taken up from the front of statistical techniques like Discriminant Analysis & T-tests, and measures like Altman Z-Index & Zeta Scores have been developed which are still in use. However, the problem makes more sense from an economic standpoint if the a successful forecast can be made about a company at least 3 years in advance. This is what has been our primary objective in this project. Having modelled this problem as a binary classification one, we have successfully achieved 80% accuracy in classification using Ensemble Machine Learning models like Random Forests, XG Boosted Tree Classifiers and Stacking Classifier. These models were built upon taking 3 years of company’s financial data and predicting outcome 3 years in advance. We concluded that long-term capital structure is a very important indicator of financial health, which we speculate, correlates with the level of capital investment required in that industry. Higher Capex is bound to lead to higher borrowings from debt. Another important factor is the Cash Conversion Cycle. Companies with regular cash flows, especially accounts payables are likely to have sound financial health. Due to dearth of data, we could not incorporate more periods into our model. As next steps, a more complex time-series trend model can be devised into the same pipeline. This might capture changing trends of different features over the years and hence, can prove to be more substantial as well as explanatory. This study can be used as building blocks of a more explainable, heuristics-based approach to decoding tendency towards going bankrupt of any given firm