Regime-aware deep learning for financial forecasting: an adaptive short-term bias compensation framework for the Warsaw Stock Exchange

Artificial Intelligence has led to significant advances in financial market, however, its application in stock market prediction is still not fully explored. This research proposes a novel Regime-Aware Adaptive Short-Term Bias Compensation (RA-STBC) framework for financial time-series forecasting. This study uses 20 years (2005–2025) of historical WIG20 index data from the Warsaw Stock Exchange (WSE) to develop a framework, which is validated across 30 independent training runs. The findings show a statistically significant 47.16% decrease in normalized Mean Squared Error (MSE) compared to Vanilla LSTM (Wilcoxon z = − 4.78, p < 0.001; Diebold-Mariano DM = 10.34, p < 0.001). A Fixed STBC ablation baseline, calibrated via validation-set grid search (λ = 0.9, uniform), achieves a larger MSE reduction of 92.27% but generates 82 trades and a −16.39% portfolio return, demonstrating that MSE minimisation alone does not translate to trading utility. The largest improvement (82.34% MSE reduction) is shown in the crisis regime, supporting the premise that adaptive correction is most beneficial when market stress is at its highest. This study’s key contribution is the integration of data-driven market regime detection with an LSTM-based forecasting model whose bias correction strength (λ) adapts dynamically to the detected regime. Unlike the prevailing approaches, the proposed framework integrates structured knowledge about market dynamics directly into the correction mechanism, which enhances the predictive accuracy of deep learning models in non-stationary financial environments.