Contrastive Learning and Self-Supervised Pre-Training for Stock Prediction: A Graph-Based Representation Learning Approach

Abstract

The scarcity of labeled financial data presents a fundamental challenge for supervised learning approaches to stock prediction, as the cost of obtaining expert-labeled data is prohibitively high while the amount of unlabeled market data grows exponentially. Self-supervised contrastive learning offers a principled solution by enabling models to learn useful representations from unlabeled data through the optimization of contrastive objectives that encourage augmented views of the same sample to be similar while views of different samples are dissimilar. This paper proposes a novel self-supervised pre-training framework for stock prediction based on contrastive learning over graph-augmented temporal views of financial time series. Our approach, Contrastive Stock State Space Graph (CS-S3G), builds upon the Stock State Space Graph architecture introduced by Lu, Hu, and Zhang, and integrates temporal and graph augmentation strategies to learn robust stock representations without relying on labeled data. We adapt explanation-based bias decoupling regularization principles from Zang and Liu's work on natural language inference to ensure that the learned representations capture genuine market dynamics rather than spurious correlations. Through extensive experiments on benchmark financial datasets, we demonstrate that self-supervised pre-training substantially improves downstream prediction accuracy, particularly in data-scarce scenarios, and that the learned representations transfer effectively across different markets and asset classes. Our work contributes to the growing body of research on self-supervised learning in finance, providing a principled framework for leveraging unlabeled data to improve prediction performance.