Prediction Targets Selection and Backtesting Performance of Machine Learning Models for Investment in Dynamic CSI 300 Constituents
Аннотация
While numerous studies have explored machine learning models for stock price prediction using various prediction targets, few studies have paid attention to the impact of target selection on investment performance. We have examined how different prediction targets affect machine learning models’ performance in stock investment using dynamic Chinese Securities Index 300 (CSI300) constituents from 2012 to 2023. We analyze six prediction targets: returns, excess returns, excess returns over risk-free rate, and their corresponding movement directions. Our research includes four machine learning models: logistic regression, multilayer perceptron, extreme gradient boosting, and transformer-based TabNet, to evaluate both predictive accuracy and investment performance through comprehensive backtesting. Results reveal that models targeting movement direction prediction significantly outperform return prediction models across key metrics including total returns, Sharpe ratios, and maximum drawdown. The movement direction of excess returns emerges as the most effective prediction target, consistently delivering superior performance across different models and portfolio sizes. In addition, equal-weight portfolio allocation strategies generally outperform weight-by-prediction-output approaches. We contribute to the literature by providing empirical evidence on the effectiveness of different prediction targets in machine learning-based investment strategies and offer practical insights for implementation in dynamic market environments.