Parameter Instability and Estimation Noise
A structural fragility of the Markov Regime-Switching Model. The parameters of a regime can only be learned from observations that fall inside that regime — with, for example, ~11 postwar US recessions, recession-state parameters rest on very thin data, and richly parameterised models become overfitted and misspecified (Hamilton). The EM likelihood also has multiple local maxima, so estimation must be restarted from many points or it converges to spurious solutions. Hess (2006) and Michaud (1989) identify noisy parameter estimates as a primary reason regime-based strategies underperform after costs — optimised portfolios behave as “estimation error maximizers.” Hamilton’s remedy is parsimony: few regimes, few switching parameters.