Stock Return Prediction: A Lightweight Innovative Framework Based on Feature Engineering and Tree Model Ensemble
TL;DR
A lightweight framework integrating simple feature engineering and XGBoost-LightGBM ensemble is proposed, by adding 3 business-logic-based dynamic features (trend, weekday cycle, volatility trend) and fusing the two tree models with equal weights.
Abstract
. Stock return prediction is a classic challenge in financial time series analysis. This study uses 996 valid daily trading samples of Netflix (NFLX) stock from 2018 to 2022, building a pure XGBoost baseline with 12 traditional technical indicators. The baseline achieves near-perfect training fit (MSE=0.0000, R²=1.000) and good test performance (MSE=0.0002, R²=0.720), but suffers from single feature dimension, weak generalization, and performance bottlenecks. To address these, the paper proposes a lightweight framework integrating simple feature engineering and XGBoost-LightGBM ensemble, by adding 3 business-logic-based dynamic features (trend, weekday cycle, volatility trend) and fusing the two tree models with equal weights (0.5:0.5). Experiments show test MSE drops to 0.000157 (3.02% reduction) and R² rises to 0.7325 (1.15% improvement). The framework is simple, stable, interpretable, and suitable for small-sample financial time series prediction.