Evaluating ensemble learning techniques for stock index trend prediction: a case of China

Abstract

Stock index trend prediction is a very important topic in the finance. The purpose of this paper is to compare six ensemble learning related techniques for stock index direction prediction, including four boosting methods (Categorical Boosting (Cat Boost), Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost) and Gradient Boosting Decision Tree (GBDT)), one bagging method (Random Forest (RF)) and one tree-structured machine learning method (Decision Tree (DT)). The Shanghai Composite Index is chosen for experimental evaluation. A factor library of seventy-two technical factors, thirty-five macro factors and seven micro factors are our inputs. Our predictions are one month ahead, and each pre diction model is evaluated by the Area Under Curve (AUC). The results indicate that ensemble learning techniques perform well in stock index prediction, with all AUC values above 0.5. RF is considered as the top algorithm with an AUC value of 0.7355 before feature selection and 0.6736 after feature selection. Also, we predict the stock index trend using a comprehensive factor library and three single factor libraries, respectively. The results show that forecasting stock index directions with a complete factor library is of great importance, which could achieve more stable forecasting results. This study contributes to literature in that it is, to the best of our knowledge, the first to make an extensive evaluation of ensemble learning related methods by constructing a comprehensive factor library and three single factor libraries.