A Fuzzy Regime-Adaptive Statistical-Deep Learning Hybrid Framework for Stock Market Forecasting

Abstract

Stock market forecasting is challenging due to the presence of nonlinearity, non-stationarity, volatility clustering, and regime uncertainty in financial time series. While statistical models effectively capture linear dependence and volatility dynamics, deep learning models are better suited for nonlinear temporal pattern extraction. However, many existing hybrid approaches do not explicitly address uncertain and regime-dependent market behavior. To overcome this limitation, this study proposes a Fuzzy Regime-Adaptive Statistical–Deep Learning Hybrid Framework for stock forecasting. The proposed framework integrates ARIMA for modeling linear structure, GARCH for capturing conditional volatility, a fuzzy inference system for uncertain market-state characterization, and a Bidirectional Long Short-Term Memory network with attention for nonlinear sequential learning. A fuzzy regimeadaptive fusion mechanism is further employed to dynamically combine statistical and deep
learning forecasts according to prevailing market conditions. The proposed framework is designed to improve predictive robustness, interpretability, and adaptability under changing financial regimes. By combining statistical rigor, fuzzy reasoning, and deep sequential learning in a unified architecture, the methodology offers a compact and theoretically grounded approach for stock market forecasting.