Interpretable Extreme Gradient Boosting Models for Multi-Class Stock Movement Prediction Using Technical Indicators

Abstract

Accurately forecasting short-horizon stock price movements remains a challenging problem in financial markets, as classical financial theory suggests that prices in liquid markets quickly incorporate available information. Nevertheless, empirical evidence continues to document short-term patterns related to momentum, volatility, and regime shifts, motivating the use of technical indicators within modern machine learning frameworks. While recent studies report promising predictive performance, many emphasize aggregate accuracy, apply limited hyperparameter tuning, or provide little interpretability, leaving uncertainty about the robustness and economic meaning of their results. The purpose of this study is to evaluate whether Extreme Gradient Boosting models built exclusively on technical indicators can reliably and transparently classify short-term stock movements into down, neutral, and up regimes. Using daily data for three large-cap U.S. equities—Tesla, Apple, and Nvidia—we construct a feature set of 36 price- and volume-based technical indicators. A two-stage hyperparameter optimization strategy combining adaptive cross-validation with refined grid search is employed, followed by strict out-of-sample evaluation using a chronological train–test split. Model interpretability is examined through global feature importance and local, instance-level explanations. The results show consistently higher sensitivity for directional movements than for neutral regimes across all assets, with momentum-based indicators, particularly Rate of Change, emerging as the dominant predictors. Local explanations reveal that strong alignment among momentum and oscillator signals underpins confident directional predictions, whereas mixed signals lead to neutral classifications. Overall, the findings demonstrate that well-tuned and interpretable models can extract economically meaningful structure from technical indicators, while also clarifying the limitations of predicting low-signal market regimes.