Bitcoin Market Regime Detection:
Hidden Markov Models Explained

The 4 Bitcoin Market Regimes

Bitcoin's price action is not random — it clusters into distinct behavioral states that quant researchers call market regimes. Each regime has a characteristic mean daily return (μ) and standard deviation (σ), making them statistically distinguishable from each other.

BTCMonitors identifies four regimes that collectively explain the full range of Bitcoin's historical price behavior:

What is a Hidden Markov Model?

A Hidden Markov Model (HMM) is a probabilistic graphical model that describes a system which transitions between a finite number of hidden states over time. The states are "hidden" because they cannot be directly observed — only the outputs (observations) generated by each state are visible.

In the context of Bitcoin trading:

• Hidden states = market regimes (Bull, Bear, HV Bear, Transition)
• Observations = daily log-returns and realized volatility
• Transition probabilities = probability of regime persisting or switching tomorrow
• Emission distributions = the Gaussian distribution of returns within each regime

An HMM makes the Markov assumption: the next state depends only on the current state, not on any history before it. This is a simplification, but it captures the essential dynamics of regime persistence and switching in financial markets.

What makes HMMs powerful for crypto trading is that they model the entire distribution of returns in each regime — not just direction. This allows them to distinguish between a "calm bull" (low σ, positive μ) and a "recovery bounce" (high σ, positive μ), which look superficially similar but have very different risk profiles.

The Mathematics of HMM

A standard HMM is defined by three parameter matrices, collectively written as λ = (A, B, π):

Transition Matrix A

The N×N matrix where A[i][j] = P(state_t = j | state_{t-1} = i). Each row sums to 1. For BTCMonitors' 4-state model:

Emission Distribution B

For a Gaussian HMM, each state emits observations from a multivariate normal distribution. For state k:

Initial State Distribution π

The vector π[i] = P(state_1 = i) gives the probability of starting in each state. For an ergodic HMM (all states reachable from all states), π is typically the stationary distribution of A.

Stationary Distribution

The long-run fraction of time spent in each regime is given by the stationary distribution π* = π* · A, solved as the left eigenvector of A corresponding to eigenvalue 1. This tells us how long Bitcoin "should" spend in each regime on average.

Baum-Welch Algorithm (EM for HMMs)

The Baum-Welch algorithm is an Expectation-Maximization (EM) algorithm that learns the HMM parameters from data without knowing the hidden state sequence. It alternates between two steps:

E-Step: Compute Forward-Backward Probabilities

The forward variable α_t(i) = P(o_1, o_2, …, o_t, state_t = i | λ) is computed recursively:

The backward variable β_t(i) = P(o_{t+1}, …, o_T | state_t = i, λ) is computed similarly from right to left.

M-Step: Re-estimate Parameters

Using the forward-backward variables, update the model parameters to maximize log-likelihood:

This process repeats until convergence (change in log-likelihood below threshold). BTCMonitors trains on 3 years of daily BTC returns and converges in ~50 iterations.

Viterbi Algorithm (State Decoding)

Once parameters are estimated, the Viterbi algorithm finds the single most likely sequence of hidden states given the observations. It's a dynamic programming algorithm that runs in O(N²T) time where N = number of states, T = sequence length.

The result q* = (q*_1, q*_2, …, q*_T) is the Viterbi path — the globally optimal regime sequence. This is what BTCMonitors displays in the "Viterbi Decoded Path" panel.

How to Trade Using Regime Signals

Regime detection is most valuable not as a directional predictor, but as a risk sizing tool. Here are the key practical applications:

1. Dynamic Position Sizing

• Low Vol Bull: Full risk allocation. Trend-following performs best. Consider 1–1.5× normal position size.
• Low Vol Bear: Reduce to 50% risk. Avoid catching falling knives. Short bias with tight stops.
• High Vol Bear: Minimum exposure (10–20%). If forced to hold, hedge with puts or stablecoins. This regime has the worst Sharpe ratio of all four states.
• Transition: 50–70% risk. Wait for the next regime to establish. Avoid momentum strategies — they underperform in mixed regimes.

2. Strategy Selection

Different strategies have different regime affinities. Trend-following excels in Bull; mean-reversion works in the low-vol phases of both Bull and Bear; options selling (volatility harvesting) is only viable in Bull; volatility buying (long gamma) is profitable in HV Bear entries.

3. Risk Management Thresholds

Use the regime's σ parameter to set dynamic stop-losses: stop = entry_price × (1 − 2.5 × σ_regime). This gives regime-aware stops rather than arbitrary percentage stops that fail to account for the current volatility environment.

4. Regime Persistence as a Filter

A regime that has persisted for fewer than 3 days has a higher probability of being a false signal. Most effective regime-based signals are conditioned on the regime being active for at least 3–5 consecutive periods — matching the expected duration implied by the diagonal of the transition matrix.

Frequently Asked Questions