Ensemble Methods and Model Diversity for Robust Stock Prediction: Combining Graph-Based State Space Models Through Bayesian Model Averaging and Stacking

Abstract

Stock prediction models face inherent uncertainty in model selection, parameter estimation, and feature engineering—a challenge compounded by the non-stationary nature of financial markets. Individual models, regardless of their sophistication, cannot capture all relevant patterns in market data, and their predictions can be overly sensitive to particular model specifications or data realizations. Ensemble methods offer a principled approach to addressing these challenges by combining predictions from multiple diverse models, reducing prediction variance, improving robustness to model misspecification, and providing natural uncertainty estimates through prediction disagreement. This paper proposes a comprehensive ensemble framework for stock prediction based on the Stock State Space Graph architecture, integrating Bayesian model averaging, stacked generalization, and bootstrap aggregation to improve prediction robustness and accuracy. Our approach, Ensemble Stock State Space Graph (E-S3G), combines multiple S3G models with diverse configurations—including different graph structures, state space dimensions, and temporal horizons—through weighted aggregation based on Bayesian posterior probabilities and meta-learner training. Through extensive experiments on benchmark financial datasets, we demonstrate that E-S3G substantially improves prediction accuracy, reduces variance, and provides well-calibrated uncertainty estimates compared to individual model baselines. Our work contributes to the growing literature on ensemble learning in quantitative finance, providing a principled framework for combining graph-based state space models to achieve robust and reliable stock prediction.