DEV Community

Cover image for Using Hidden Markov Models for Crypto Regime Detection
Gunnar Thorderson
Gunnar Thorderson

Posted on • Originally published at getregime.com

Using Hidden Markov Models for Crypto Regime Detection

#machinelearning #crypto #hmm #python

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',
    }
Enter fullscreen mode Exit fullscreen mode

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 ]
Enter fullscreen mode Exit fullscreen mode

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:

  1. Weighted ensemble → primary regime classification (what the API returns)
  2. 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
Enter fullscreen mode Exit fullscreen mode

Returns:

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

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

  1. HMMs detect transitions earlier but have more false positives in steady state
  2. Bear states are sticky — once bear, 93.5% chance of staying bear
  3. Bull "detections" are often bounces — the HMM finds temporary rallies within larger bear trends
  4. 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

Try Regime Intelligence

Regime is a real-time crypto market regime detection API. One endpoint tells you if the market is bull, bear, or chop — so your bot only trades when conditions match your strategy.

Free API access → | See pricing → | API docs →

Top comments (0)