Comparative performance of machine learning-selected portfolios from dynamic CSI300 constituents: forward vs. backward adjusted stock prices

Abstract

Most existing studies utilize backward-adjusted stock prices from data platforms to develop and backtest investment strategies using machine learning models. However, these prices are not point-in-time data and may introduce look-ahead bias, raising concerns about the reliability of model performance. To examine the impact of different price adjustment methods, we compare the predictive performance of various machine learning models and the backtesting results of portfolios constructed using these models with both forward-adjusted and backward-adjusted stock prices. Our study, conducted from 2012 to 2022, evaluates the real-world viability of investment strategies on the dynamic constituents of the CSI300 index. The empirical results reveal that while certain measures of machine learning models’ predictive performance may not be significantly affected by the stock price adjustment method, the backtesting performance under backward-adjusted stock prices is overestimated compared to that under forward-adjusted stock prices. This research provides evidence for the impact of historical stock price adjustments in developing machine learning models and presents a comprehensive framework for applying these techniques to the management of index constituent portfolios, thereby bridging the gap between predictive modeling and practical investment strategies.