Dynamic feature selection with distributional Markov chain error correction for financial time series forecasting

Financial time series forecasting presents fundamental challenges requiring adaptive methodologies that can handle uncertainty, high-dimensional feature spaces, and regime-dependent market dynamics. This paper introduces a novel adaptive framework that integrates dynamic threshold-based feature selection with distributional Markov chain error correction for enhanced financial forecasting performance. The proposed approach combines Adaptive Threshold-based Feature Selection with theoretically-grounded bounds that dynamically adjust to evolving market characteristics, and Markov Chain with Sample Distribution that captures both regime transitions and intra-state uncertainty through distributional modeling. Empirical validation on S&P 500 data demonstrates consistent improvements over traditional econometric and machine learning baselines, with statistically significant gains across multiple evaluation criteria. Robustness analysis across different market regimes confirms the method’s stability and effectiveness for financial forecasting. The framework’s principles of adaptability, uncertainty handling, and hybrid modeling make it particularly suitable for financial markets requiring intelligent prediction under uncertainty.