A temporal graph-based contrastive approach for financial time series forecasting

Graph neural networks have been used for semi-supervised learning with inter-related stocks, but their application to capturing temporal patterns in financial time series is limited. Additionally, deep similarity-based learning, which is effective for learning distinctive embeddings and overcoming nonlinearity, has not been extensively studied in financial time series forecasting. In this study, we propose a temporal graph-based contrastive approach for forecasting multi-step movement direction in stock markets. We introduce a method for generating contrastive samples based on market behavior, followed by representing time series fluctuations using graphs. Then, we propose a graph-based contrastive learning approach to embed graph samples into a new space, bringing similar samples closer and distancing dissimilar ones. Ultimately, predicting movement direction is simplified to neighborhood analysis. The proposed framework outperforms state-of-the-art benchmarks across nine major market indices, delivering improvements in Balanced Accuracy (1%–5%) and F-score (1%–6%) over various horizons. It also demonstrates excellence in financial metrics such as the Sharpe Ratio and Maximum Drawdown, emphasizing both predictive performance and practical applicability in financial analysis.