Ensembling and Dynamic Asset Selection for Risk-Controlled Statistical Arbitrage

In recent years, machine learning algorithms have been successfully employed to leverage the potential of identifying hidden patterns financial market behavior and, consequently, become a land opportunities for applications such as algorithmic trading. this paper, we propose statistical arbitrage trading strategy with two key elements: an ensemble regression asset return prediction, followed by dynamic selection. More specifically, construct extremely heterogeneous ensuring model diversity using state-of-the-art algorithms, data feature selection process, and method individual models each asset, well that learn cross-sectional across multiple assets. Then, their predictive results are fed into quality assurance mechanism prunes assets poor forecasting performance in previous periods. We evaluate approach on historical component stocks S&P500 index. By performing in-depth risk-return analysis, show setup outperforms highly competitive strategies considered baselines. Experimentally, enhances overall both terms risk. Moreover, proposed proved yield superior during turmoil massive growth periods, it showed general application any risk-balanced aiming exploit different classes.