Wavelet-enhanced multimodel framework for stock market forecasting: A comprehensive analysis across market regimes

In this study, we develop a hybrid forecasting framework that integrates discrete wavelet transform with multiple machine and deep learning architectures to address nonlinearity and regime-dependent dynamics in financial markets. Log-return series using daily data from the BIST 100, S&P 500, and Shanghai Composite indexes spanning 2015–2025 are subject to three-level Daubechies-4 wavelet decomposition, which yields approximation and detail coefficients that capture multiresolution temporal patterns. Four feature configurations are systematically evaluated: base (lagged returns only), pure wavelet approximation (A1–A3), hybrid wavelet approximation (lags combined with A1–A3), and wavelet approximation-detail (A1–A3 with D1–D3). Random forest, Support Vector Regression, Long Short-Term Memory, and Gated Recurrent Unit models are trained on each configuration, enabling direct assessment of the effectiveness of wavelet feature engineering. A three-state Gaussian hidden Markov model identifies bull, bear, and sideways regimes based on risk-adjusted returns, stratifying out-of-sample results to examine model robustness across varying market conditions without influencing training procedures. Our results demonstrate that wavelet-enhanced configurations, in particular the full approximation-detail specification, reduce forecast errors by 20–40 percent across all indexes and algorithms. Diebold–Mariano tests confirm statistical significance both globally and within each market regime. Our findings confirm that discrete wavelet transform is essential preprocessing for volatile financial markets, offering actionable insights for algorithmic trading, risk management, and policy frameworks in emerging economies.