HMM Parameter Instability

HMM parameter instability is the failure mode in which the parameters of a Hidden Markov Model Regime Detection — the transition matrix and the per-state emission distributions — are estimated noisily and unreliably from limited financial data. It is the model-specific instance, for the hidden-state Markov family, of the general Parameter Instability and Estimation Noise failure mode. Three distinct mechanisms drive it, and they compound.

First, data starvation of minority regimes. The bear / high-volatility regime occupies only a small fraction of any equity history, so the observations available to estimate its emission mean and variance are few. Shu, Yu & Mulvey (2024) name mis-estimation “from limited sample sizes, unbalanced data and high state persistence” as a core HMM weakness — the persistence that makes regimes useful also means each regime contributes few independent transitions to its own parameter estimates. Hamilton’s remedy in the broader regime-switching literature is the same conclusion reached from the other direction: with only a handful of postwar US recessions, recession-state parameters rest on very thin data, and the answer is parsimony — few regimes, few switching parameters.

HMM Parameter Instability [part-of] Parameter Instability and Estimation Noise HMM Parameter Instability [causes] Regime Misclassification

Second, multi-modal likelihood and local optima. HMM parameters are fitted by Baum-Welch Estimation, the EM specialisation for HMMs. Baum-Welch guarantees only a non-decreasing likelihood and converges to a local optimum, not the global maximum; finding the globally optimal parameter set is NP-complete. The likelihood surface is genuinely multi-modal — an algorithm-comparison study (arXiv:2409.02477) finds “the presence of multiple optima… seems to play a role” in estimator behaviour, with quasi-Newton variants converging to local optima more often than Baum-Welch. The practical consequence is that the fit depends on the random starting point: the standard defence is to run many random initialisations and keep the highest-likelihood result, “without formal guarantees on escaping all local optima.” A fitted HMM is therefore not a single deterministic object — a different seed can yield materially different regimes. Baum-Welch is additionally numerically unstable in naive form, because it recursively multiplies joint probabilities that underflow toward zero over long sequences, requiring scaling or log-space arithmetic.

HMM Parameter Instability [causes] Baum-Welch Estimation

Third, rolling-window drift. In any online regime strategy the HMM is re-fitted as the estimation window moves, so the parameters change period to period even absent a true regime shift. This couples HMM parameter instability to the Non-Stationary Transition Matrix problem: part of the apparent regime activity is real non-stationarity, part is estimation noise re-sampled each window, and the two are hard to separate.

The trading cost of all this is concrete. Noisy parameters produce short-lived, spurious regimes — the model flips state on what is really estimation noise — and each flip is a costed trade. This is the upstream cause of Regime Misclassification, whose error rates spike to ~10% at window edges, and it is part of why the HMM-guided strategy in Shu Yu and Mulvey 2024 ran turnover of 141-290% against the Statistical Jump Model’s 44-72%, eroding its net edge. For this vault’s purpose the verdict is a confirmed structural fragility: HMM parameter instability is not a tuning nuisance but a mechanical reason regime-timing backtests are hard to reproduce and lose money to turnover after costs.

HMM Parameter Instability [relates] Non-Stationary Transition Matrix Shu Yu and Mulvey 2024 [supports] HMM Parameter Instability

Connections

Sources