Using Hidden Markov Models for Crypto Regime Detection

Most market regime classifiers use simple threshold rules: SMA above 200 = bull, below = bear. These work but they're noisy and laggy. Hidden Markov Models (HMMs) offer a probabilistic alternative that can detect regime transitions 6-12 hours earlier.

Here's how we built one.

Why HMMs for Regime Detection

Markets transition between regimes (bull/bear/chop) without announcing it. You can observe symptoms — volatility changes, funding rate shifts, sentiment swings — but the underlying regime is hidden. This is exactly what HMMs are designed for: inferring hidden states from observable emissions.

Architecture

Our HMM has:

3 hidden states: Bull, Bear, Chop
6 observations: SMA ratio, volatility ratio, funding rate, fear/greed, volume level, BTC dominance
Training: Baum-Welch algorithm on 3,291+ market snapshots
Inference: Viterbi algorithm for most likely state sequence

The Observations

Each observation is discretized from continuous market data:

# Simplified observation generation
def make_observation(snapshot):
    return {
        'sma_ratio': 'bullish' if snapshot.sma50 > snapshot.sma200 else 'bearish',
        'volatility': 'high' if snapshot.atr20/snapshot.atr90 > 1.3 else 'low',
        'funding': 'positive' if snapshot.funding_rate > 0.01 else 'negative' if snapshot.funding_rate < -0.01 else 'neutral',
        'fear_greed': 'greed' if snapshot.fg > 60 else 'fear' if snapshot.fg < 40 else 'neutral',
        'volume': 'high' if snapshot.volume_ratio > 1.2 else 'low',
        'dominance': 'rising' if snapshot.btc_dom_delta > 0.5 else 'falling' if snapshot.btc_dom_delta < -0.5 else 'flat',
    }

Training with Baum-Welch

The Baum-Welch algorithm (a special case of EM) estimates the transition and emission probabilities:

Transition matrix (what we learned):
         → Bull   → Bear   → Chop
Bull:    [ 0.017   0.821   0.162 ]
Bear:    [ 0.012   0.935   0.053 ]
Chop:    [ 0.089   0.456   0.455 ]

Key finding: Bear states are extremely sticky (93.5% self-transition probability). Once the market enters a bear regime, it tends to stay there. Bull detections are transient (1.7% stay probability) — the HMM is detecting bounces within bear markets rather than true regime shifts.

Results vs Weighted Classifier

We compared the HMM against our 10-signal weighted ensemble:

Metric	HMM	Weighted Ensemble
Agreement rate	68.4%	—
Bear detection	Earlier by 6-12h	More stable
Bull detection	Noisy (false positives)	More reliable
Chop detection	Poor (rare state)	Better (threshold-based)

The HMM excels at detecting transitions early. The weighted ensemble excels at steady-state classification. We use both:

Weighted ensemble → primary regime classification (what the API returns)

HMM → early warning signal for upcoming transitions (available in intelligence brief)

Accessing the HMM

The HMM regime is available via the Regime API (Pro tier):

curl -H "Authorization: Bearer YOUR_KEY" \
  https://getregime.com/api/v1/intelligence/hmm-regime

Returns:

{
  "hmmRegime": "bear",
  "hmmConfidence": 0.89,
  "weightedRegime": "bear",
  "agreement": true,
  "transitionProbability": 0.065,
  "lastRetrained": "2026-03-26T04:00:00Z"
}

When agreement is false, a regime transition may be imminent. The transitionProbability indicates how likely the current regime is to change in the next evaluation window.

Key Takeaways

HMMs detect transitions earlier but have more false positives in steady state

Bear states are sticky — once bear, 93.5% chance of staying bear

Bull "detections" are often bounces — the HMM finds temporary rallies within larger bear trends

Combining HMM + threshold classifier gives the best of both worlds

Free regime endpoint (weighted ensemble): curl https://getregime.com/api/v1/market/regime

Full intelligence including HMM: Pro tier

GitHub: getregime.com