Abstract
Long‐term corporate solvency forecasting is fundamentally constrained by two intertwined obstacles: severe class imbalance in default events and the nonstationary dynamics of financial markets. Deep learning architectures optimized on short‐horizon objectives characteristically overfit to transient market noise, yielding predictive performance that deteriorates sharply as the forecasting horizon extends. This instability is addressed through the Gated Spectral–Temporal Transformer (GST‐Net) a hybrid architecture that introduces a learnable gating mechanism to suppress transient, high‐frequency signal components in favor of persistent structural trends. A dual‐branch design further decouples the modeling of localized anomalies from global temporal dependencies, enabling the network to preserve discriminative signal across multiyear horizons where conventional models degrade. Given the extreme class imbalance inherent to solvency prediction, the evaluation protocol centers on the Matthews correlation coefficient and macro F1 score, metrics that penalize minority‐class misclassification far more rigorously than accuracy alone. At a three‐year forecasting horizon, GST‐Net achieves an MCC of 0.5704, a macro F1 score of 78.38%, and an accuracy of 95.95%, nearly doubling the MCC of the strongest tuned baseline, XGBoost, which peaked at 0.3294, while state‐of‐the‐art tabular deep learning models, TabNet and TabPFN, collapse entirely under identical conditions. Beyond predictive robustness, GST‐Net sustains subsecond inference latency of approximately 0.37 s per instance, rendering it practical for quarterly operational auditing at scale. Coupling the model with SHAP and LIME yields the interpretability required for regulatory scrutiny, exposing an economically coherent shift from short‐term liquidity stress to long‐term structural decay as the dominant predictive signal, an insight with direct implications for early‐warning credit risk systems.