Boosting Algorithms to Analyse Firm’s Performance Based on Return on Equity: An Explanatory Study

Abstract

This study aims to use the boosting algorithms especially gradient boosting and its extension extreme gradient boosting in predicting firm performance in terms of return on equity that could be considered as a measure of profitability and to use the partial dependent plot and local interpretable model-agnostic explanation techniques to explain the model and its prediction. The models are evaluated using R-squared, root mean square error, and mean absolute error. The global interpretations in terms of partial dependent plot and local interpretations in terms of local interpretable model-agnostic explanations are performed to interpret the prediction for any individual or group of cases. The results show that the extreme gradient boosting is improving the model by about 39% for training set and about 4% for testing set in terms of R-squared. Interesting results are given by the partial dependent and local model-agnostic explanation plots where they are suggesting that the total assets, the total liability and the board size have the most effect on the predicting and interpreting return on equity. By taking over-fitting in consideration the gradient boosting model is a better choice than extreme gradient boosting. The important scores suggest that the total liability, beta coefficients and the total assets have the most effect on return on equity.