CalixBoost: A Stock Market Index Predictor using Gradient Boosting Machines Ensemble

Abstract

The potential of machine learning has sustained the interest of both academia and industry in stock market prediction over the past decade. This paper aims to integrate modern techniques such as Gradient Boosting Machines (GBMs) into a novel ensemble called CalixBoost which is a resource-efficient and accurate stock index predictor. Data comprising macro-economic metrics and technical financial indicators, as well as sentiment analysis of social media using a simple and fast but effective rule-based model are used in this paper. Other techniques include model tuning with Bayesian Optimization, temporal consistency analysis for invariant feature selection over random trial-and-error, feature importance and inter-feature relationships analysis using a unified game theory approach using Shapley values. Lastly, the model will be evaluated using a novel holdout method, viz. on two separate test datasets whose datapoints are collected under (i) normal economic activity and (ii) during a black swan (financial downturn). The experimental results show that our model outperforms previous methods and can achieve a good prediction performance with 84.88% accuracy, 0.0956 RMSE, 0.0573 MAE and 4.19% MAPE.