Predicting stock splits using ensemble machine learning and SMOTE oversampling

This study predicts stock splits using two ensemble machine learning techniques: gradient boosting machines (GBMs) and random forests (RFs). The goal is to form implementable portfolios based on positive predictions to generate abnormal returns. Since splits are rare events, we use SMOTE oversampling to synthesize new observations of splits in the sample to improve predictions. When predicting stock splits in the next quarter, GBM and RF achieve area under the receiver operating characteristic curve (AUC) scores of around 0.86 and 0.87, respectively. GBM and RF predictions generate monthly five-factor alphas (Fama and French, 2015) of 0.26% and 0.95% among stocks in the smallest size quintile. Three important features for predicting stock splits in both ensemble ML methods are current price levels, the ratio of current price to the price at last split, and stock returns in the past twelve months. When predicting stock splits in the next year, GBMs generate monthly five-factor alphas of 0.38% among small stocks.