TRADE-OFF BETWEEN EXPLAINABILITY AND PERFORMANCE IN ASSET PRICING MODELS

Abstract

In recent years, the Indian stock market has flourished due to enthusiastic investor participation and robust secondary market infrastructure. The technological advancements in the market have led to the adoption of sophisticated models for accurate stock return predictions. Machine Learning (ML) algorithms are playing a significant role in this process by capturing complex patterns and infusing non-linearity to classic asset pricing models. However, many advanced ML techniques are “black box” in nature, lacking transparency in their workings. This reduces the interpretability of such models, which is integral for practitioners in finance, such as mutual fund managers, portfolio managers, institutional investors, retail investors and traders. In this Thesis, the trade-off between performance and explainability of asset pricing models has been investigated by exploring various ML Techniques (Regression, ANN, SVM, Ridge Regression, Panel Regression and Panel ANN) and comparing them based on their performance metrics (R2, RMSE, MASE, MAE and MedAE) and explainability metrics (coefficients or permutation importance) to find the appropriate model. ANN has been found to be significantly better than other models based on performance metrics. A survey has also been conducted to gain insight into people’s perceptions regarding the trade-off between the performance and explainability of ML models by identifying the threshold point at which performance outweighs explainability.