Machine Learning for Explainable Future Earnings Forecasting

Abstract

This thesis explores the application of machine learning (ML) to forecast the direction of future quarterly earnings. It leverages high-dimensional data from financial statements and applies the Light Gradient Boosting Machine (LightGBM) algorithm. It introduces a feature transformation approach that incorporates bi-quarterly and tri-quarterly changes, which improves the model’s accuracy. The findings show that this ML model closely matches the performance of analysts in predicting earnings direction. More importantly, a trading strategy based on the model’s predictions outperforms one based on analysts’ forecasts, underscoring the model’s economic significance. This model also achieves higher classification accuracy when benchmarked against two approaches: the one followed in [22] and a parametric estimation method. Additionally, the thesis applies Explainable Artificial Intelligence (XAI) techniques, specifically SHAP and LIME, to provide both global and local interpretations of the model’s predictions. This provides insights into how different financial features influence the model’s decisions. Overall, this study makes two significant contributions to earnings forecasting literature. First, it establishes that a ML model predicting the direction of quarterly earnings, a previously unexplored prediction approach, can be both accurate and economically significant. Second, it demonstrates the use of state-of-the-art XAI techniques in explaining the decision-making process of ML models in the context of forecasting changes in earnings.

Date of Award	1 May 2024
Original language	American English
Supervisor	PANAGIOTIS Liatsis (Supervisor)