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

Battle Count: 133

Highly relevant to quantitative trading in several dimensions: (1) Provides a principled framework for adaptive signal extraction that accounts for regime-dependent parameter instability; (2) Documents that foreign investor predictive power increases 3.51-fold during crisis periods, directly relevant for crisis-period trading strategies; (3) The asymmetric response analysis reveals that retail investors chase momentum 6.3x more strongly on rallies than declines, exploitable as contrarian signals; (4) The 'All-Weather' strategy framework with regime-conditional position sizing is directly applicable to portfolio construction. However, the decisive OOS failure (Sharpe = -1.65) and negative in-sample returns for most strategies temper practical applicability. The paper is more valuable as a methodological framework and descriptive characterization than as a ready-to-deploy trading system.

The integrated framework combines multiple sophisticated components: (1) Adaptive Kalman filter with heteroskedastic noise coupling to realized volatility requires careful parameter estimation (phi, Q, R0, gamma); (2) Three-state Markov-switching model with regime-conditional regressions involves EM algorithm estimation; (3) Asymmetric response functions with threshold-based indicator variables; (4) JSD-based adaptive signal blending with rolling window computations; (5) Viterbi decoding via two-state Gaussian HMM; (6) The 'All-Weather' strategy integrates all components with regime-conditional position sizing and asymmetric stop-loss rules. Each component is individually well-understood in the literature, but their joint implementation with proper parameter estimation, regime identification, and out-of-sample validation requires significant engineering effort. The paper does not provide code, adding to implementation burden.

The paper provides detailed mathematical formulations for all components (Kalman filter equations, Markov-switching specification, asymmetric response functions, JSD computation, Viterbi decoding). Data source is identified (Korea Exchange daily transaction data 2020-2024). However, no code repository is mentioned, and the specific parameter choices (e.g., Kalman filter parameters phi, Q, R0, gamma) and their estimation procedures could be more transparent. The out-of-sample validation methodology is clearly described with frozen parameters at end-2022.