Heterogeneous Ensemble Learning Model for Time Series Processing Based on Multi-Dimensional Indicator Feature Engineering

This paper proposes a heterogeneous ensemble learning model based on multi-dimensional indicator feature engineering, aiming to improve the accuracy and stability of time series predictions. First, by constructing features from 14 indicator groups, including momentum, trend, volatility, etc., the model comprehensively analyzes time series behavior. Then, for different models (XGBoost, Random Forest, Naive Bayes, and LSTM), various feature selection methods are employed to identify the features that contribute the most to predicting future. Next, multiple model training methods are applied to train the models, and a meta-learner is used to combine the predictions from multiple base models, further enhancing prediction performance. Following that, the paper optimizes the threshold in the classification task based on the existing ensemble learning model to improve classification accuracy and F1-score. Finally, the model's effectiveness is verified through backtesting on multiple real time series datasets. The results show that the heterogeneous ensemble learning prediction model demonstrates strong performance in forecasting and has significant application potential.