When the Rules Change: Adaptive Signal Extraction via Kalman Filtering and Markov-Switching Regimes

Static linear models of order flow assume constant parameters, failing precisely when they are needed most: during periods of market stress and structural change. This paper proposes a dynamic, state-dependent framework for order flow signal extraction that adapts to shifting market conditions in the Korean stock market. Using daily transaction data from 2020--2024 covering 2,439 stocks and 2.79 million stock-day observations, we implement three complementary methodologies: (1) an Adaptive Kalman Filter where measurement noise variance is explicitly coupled to market volatility; (2) a three-state Markov-Switching model identifying Bull, Normal, and Crisis regimes; and (3) an Asymmetric Response Function capturing differential investor reactions to positive versus negative shocks. We find that foreign investor predictive power increases 8.9-fold during crisis periods relative to bull markets ($\beta_{crisis}=0.00204$ vs. $\beta_{bull}=0.00023$), while individual investors exhibit momentum-chasing behavior with 6.3 times stronger response to positive shocks. The integrated ``All-Weather'' strategy provides modest drawdown reduction during extreme market events, though challenges remain in the post-COVID high-rate environment.