Model interpretability of financial fraud detection by group SHAP

Abstract

With the emergence of artificial intelligence, related technologies have gradually been deployed for the management of corporate financial risk, which is one of the utmost concerns for financial institutions. Especially, it appears from the previous attempts that the artificial intelligence models for financial risk detection show a decent performance on fraudulent firm recall. However, these models, as a black box prediction, lack the interpretability to gather clues for specific needs, like due diligence. Here we investigate the interpretability of risk detection models, which allows the contribution collection of each feature or feature group. Shapley value is set as the criterion to evaluate the marginal contribution of each financial feature to the predicted result. Moreover, the group SHAP method is proposed to evaluate the different abilities of a company, such as profitability, liquidity, etc. The group SHAP method presents an excellent performance in local interpretation and significantly decreases computation time compared with the Shapley value. It is worth mentioning that common characteristics within the same industries and uniqueness between different ones have been observed through group SHAP. With the analysis of these clues collected by the group SHAP method, financial institutes may provide instructions on focal points of due diligence for investment bankers or relevant risk alerts for risk managers while monitoring a stock within a portfolio, which has important implications for clients monitoring or investment management.