Integrating Multi-Temporal Feature Ranking with Forecasting Models for Enhanced Stock Price Prediction

Achieving high prediction accuracy remains challenging due to the inherent volatility and complexity of financial markets. This study aims to bridge the gap by integrating effective feature selection and advanced modeling, combining multi-temporal feature ranking with 10 state-of-the-art (SOTA) predictive models. By analyzing constituent stocks of the Chinese CSI 800 and CSI 1000 indices from January 4th, 2007, to September 28th, 2023, and incorporating 13 features from three temporal frequencies, we found that Bayesian LightGBM has a 5.7% higher Mean Squared Error (MSE) but a 34.8% lower Mean Absolute Error (MAE) compared to Bayesian XGBoost. A backtest on the portfolio was conducted to validate the reconstructed data. Additionally, in predictive inference, Autoformer shows an 18.68% lower Root Mean Squared Error (RMSE) and a 16.88% higher MAE than Informer. This paper offers valuable insights into the financial machine learning process and its application across various temporal datasets.