Continuous Learning and Online Adaptation for Non-Stationary Stock Markets: Addressing Concept Drift in Graph-Based State Space Models

Abstract

Financial markets are inherently non-stationary systems where the statistical relationships between predictive features and stock returns evolve continuously over time. This phenomenon, known as concept drift, poses a fundamental challenge for machine learning models trained on historical data, as patterns that held during training may no longer apply during deployment. The Stock State Space Graph model introduced by Lu, Hu, and Zhang achieves impressive prediction accuracy by integrating state space modeling with graph-based relational reasoning, yet it is designed for stationary settings where model parameters remain fixed after training. This paper proposes Adaptive Stock State Space Graph (A-S3G), a framework that extends the S3G architecture through continuous learning mechanisms that detect and adapt to concept drift in real-time. Our approach integrates concept drift detection using statistical monitoring and change point detection, online learning algorithms for incremental model updates, and elastic weight consolidation to prevent catastrophic forgetting of previously learned patterns. Through extensive experiments on benchmark financial datasets spanning multiple market cycles, we demonstrate that A-S3G substantially outperforms static S3G baselines in non-stationary environments, maintaining predictive accuracy during market regime transitions while preserving performance on previously learned patterns. Our work contributes to the growing literature on continuous learning in finance, providing a principled framework for deploying adaptive prediction models in dynamic market environments.