Prediction of Stock Returns Using Accounting Data with a Machine Learning Approach

Abstract

The relationship between accounting data and stock price prediction has been a hot topic for over half a century. Researchers have been trying to identify the relationship and investigate how it may be useful when trying to improve prediction accuracy. The non-linear relationship and unpredictable stock market environment translate to a complex forecast and prediction procedure. However, recent developments in statistics and machine learning allows for earlier technical limitations to be solved. It has been argued that machine learning models can assist in identifying and translating patterns that previously were not comprehensible. This study tests this statement by utilizing the traditional logistic regression along with a newly introduced machine learning library called CatBoost, based on the gradient boosting decision tree algorithm. This study provides evidence of the usefulness of the two models and how they improve the prediction accuracy of directional stock price movements. In addition, the relevance of using accounting data for prediction purposes is supported by the results of the study. Further, the predictive capability of individual performance measures is presented where risk and growth proxies together with profitability proxies are identified as the most important and influential predictor variables.