Financial risk forecasting with RGCT-prerisk: a relational graph and cross-temporal contrastive pretraining framework

Abstract

Financial risk forecasting is critical for the early detection of corporate distress, yet traditional methods and recent deep learning models exhibit notable limitations. Prior approaches often rely on predefined financial ratios or brute-force feature combinations, which may overlook the relational structures among financial items and suffer from high dimensionality. In this paper, we propose RGCT-PreRisk, a novel framework that models a company’s financial statements as a heterogeneous relational graph-where nodes represent financial accounts (e.g., assets, liabilities) and edges encode known accounting relationships (e.g., summation or ratio rules)-rather than as an unstructured feature matrix. A graph neural network (GNN) is employed to capture meaningful relationships between financial items (e.g., assets, liabilities, revenues), replacing exhaustive pairwise operations with learned propagation along true accounting dependencies. To address the challenge of limited labeled data, we introduce a cross-temporal, cross-company contrastive pretraining strategy that leverages historical data across multiple firms and time periods to learn robust and generalizable representations. Furthermore, we incorporate a prototype-attention-confidence module to enhance interpretability. This component enables the model to compare each firm’s financial state to learned prototypical risk patterns via an attention mechanism, while also producing a confidence score to quantify prediction uncertainty. Experiments on two real-world datasets demonstrate that RGCT-PreRisk consistently outperforms existing baselines in terms of accuracy and F1 score. Our approach achieves state-of-the-art predictive performance while providing human-interpretable insights into why a firm is predicted to be at risk. This work presents a new direction for interpretable financial risk forecasting by integrating graph-based representation learning, contrastive pretraining, and case-based reasoning.