Charlie Axelbaumfinanceresearch

Regime Detection with Hidden Markov Models

Markets move through distinct regimes, but they don't announce transitions. HMMs offer a probabilistic framework for identifying where we are — and where we might be going.

February 20, 20259 min read

Why Regime Detection Matters

The most dangerous assumption in quantitative finance is stationarity — the idea that the statistical properties of a series don't change over time. Markets aren't stationary. They shift between regimes: bull markets, bear markets, stress periods, dislocations. A model calibrated to one regime will fail in another.

Regime detection is the attempt to identify, in real time, which regime we're in. If you can do that reliably, your models become regime-conditional and dramatically more accurate.

The HMM Approach

A Hidden Markov Model treats the regime as a latent (unobserved) variable that evolves over time according to a transition matrix. The observed data — spread levels, volatility, cross-asset signals — are assumed to be generated by the current hidden state.

The model learns:

The emission distributions (what each regime "looks like" in the data)

The transition probabilities (how likely we are to stay in or leave each regime)

At inference time, we can compute the posterior probability of being in each regime at each point in time — and use the transition probabilities to forecast the next regime.

Implementation Notes

I trained the model on daily credit spread data (IG CDX, HY CDX) from 2005–2024. Features: spread level, spread vol (21-day rolling), equity implied vol (VIX), rates vol (MOVE index), and a flow proxy based on ETF data.

Key implementation choices:

3-state model (risk-on, risk-off, stress) performed better than 2 or 4

Gaussian emissions with full covariance — mixture models didn't add enough to justify complexity

Viterbi decoding for most likely path; forward-backward for smooth state probabilities

Results

The model identifies regime transitions — defined as a sustained move from one dominant state to another — with a median lead of 8 trading days compared to naive threshold-based approaches. It's particularly useful for timing systematic duration hedges in credit portfolios.

False positive rate is manageable but nonzero. The model tends to generate spurious stress signals during low-vol periods when spreads are compressing erratically. A vol filter helps.

Code

from hmmlearn import hmm

import numpy as np

model = hmm.GaussianHMM(

n_components=3,

covariance_type="full",

n_iter=200,

random_state=42

)

model.fit(features)

states = model.predict(features)

state_probs = model.predict_proba(features)

What I'd Do Differently

The main limitation is sample size. Credit cycle data is sparse — we only have a handful of full cycles. I'd want to augment with international spread data and potentially synthetic data from structural credit models to get a better-calibrated stress regime.

The other thing: don't overfit to the HMM framework. The value is the regime-conditional modeling, not the HMM itself. If a simpler clustering approach gives you cleaner regime labels, use that.