Research

Bitcoin Volatility Forecasting: Models That Actually Work

RegimeRisk · 8 May 2026 · 9 min read

VolatilityQuantitative AnalysisBitcoinRisk Models

Bitcoin Volatility Forecasting: Models That Actually Work

Bitcoin volatility forecasting is one of the harder problems in quantitative finance. Unlike equities, BTC trades 24/7 across fragmented venues, responds to macro shocks and on-chain events simultaneously, and can move 10% in an afternoon on a single news catalyst. As of early May 2026, BTC is trading around $79,880 after briefly touching $81,500 — a level not seen since January — lifted by tokenization momentum and macro tailwinds. That kind of move, roughly 8–10% in a short window, is exactly the environment where a good volatility forecast separates disciplined risk management from guesswork.

This post surveys the main approaches to bitcoin volatility forecasting, explains what each model is actually doing under the hood, and connects forecast outputs to the regime classification process that makes those forecasts actionable.

Why Volatility Forecasting Is Different for Bitcoin

Traditional volatility models were built for assets with defined trading hours, deep liquidity, and relatively stable microstructure. Bitcoin violates most of those assumptions. Key structural features that complicate crypto volatility prediction:

Continuous trading: No overnight gap, but also no natural break in the return series. Weekends are fully included. This means a 30-day window for a stock contains ~22 observations; for BTC it contains 30.
Regime clustering: Bitcoin has historically alternated between extended low-volatility accumulation phases and violent expansion phases. These aren't random walks — they cluster in ways that matter for model selection.
Liquidity fragmentation: Bid-ask spreads and order book depth vary enormously across exchanges, which contaminates high-frequency volatility estimates.
Reflexivity: Volatility itself drives news coverage, which drives retail participation, which drives more volatility. Feedback loops are tighter than in traditional markets.

None of this makes forecasting impossible. It does mean you need to choose models that can handle fat tails, structural breaks, and non-stationarity.

The Core Approaches

1. GARCH and Its Variants

GARCH (Generalized Autoregressive Conditional Heteroskedasticity) remains the workhorse of volatility modelling across asset classes, and GARCH bitcoin applications have a solid empirical track record. The intuition is simple: today's variance is a weighted combination of yesterday's squared return (the shock component) and yesterday's estimated variance (the persistence component).

The standard GARCH(1,1) specification:

σ²_t = ω + α·ε²_(t-1) + β·σ²_(t-1)

For Bitcoin, fitted GARCH models typically show high persistence (β close to 0.9 or above), meaning volatility shocks decay slowly. This matches intuition: once BTC becomes choppy, it tends to stay choppy for days or weeks.

Where standard GARCH falls short for crypto:

Symmetric response: GARCH treats positive and negative shocks identically. Bitcoin, like equities, shows some asymmetry — downside moves tend to spike implied volatility more than equivalent upside moves, though the effect is less pronounced than in equity markets.
Fat tails: Standard GARCH with Gaussian errors underestimates tail risk. Using Student-t or GED (Generalized Error Distribution) error terms significantly improves fit.
Structural breaks: A model calibrated in a low-vol accumulation phase will misforecast when a new expansion phase begins.

Variants worth knowing:

EGARCH (Exponential GARCH): Allows asymmetric volatility response and doesn't require non-negativity constraints.
GJR-GARCH: Adds a separate coefficient for negative shocks. Often outperforms standard GARCH on daily BTC returns.
IGARCH: Imposes unit persistence, which can be appropriate during sustained high-volatility regimes.

For practical bitcoin volatility forecasting at the daily horizon, GJR-GARCH with Student-t errors is a reasonable baseline.

2. Realized Volatility Models

Realized volatility (RV) uses intraday returns to construct a more precise variance estimate than end-of-day data alone. The intuition: if you have hourly returns, you can estimate daily variance much more accurately than from a single daily close-to-close return.

The Heterogeneous Autoregressive model (HAR-RV), developed by Corsi (2009), has become a standard in the RV literature. It models daily realized variance as a function of daily, weekly, and monthly realized variance components:

RV_t = c + β_d·RV_(t-1) + β_w·RV_(t-5:t-1) + β_m·RV_(t-22:t-1) + ε_t

This three-component structure captures the empirical finding that volatility is persistent across multiple time horizons simultaneously — a feature that single-lag GARCH models miss.

For Bitcoin, HAR-RV models have shown competitive out-of-sample performance at short-to-medium horizons (1–5 days). Extensions that incorporate signed jumps (separating large positive from large negative price moves) and semi-variance measures tend to improve forecast accuracy further.

The limitation: HAR-RV requires clean, high-frequency data. For Bitcoin, this means choosing a reference exchange or constructing a volume-weighted composite price — both of which introduce methodological choices that affect results.

3. Implied Volatility and Options-Derived Measures

Where liquid options markets exist, implied volatility (IV) aggregates the market's collective forecast. For Bitcoin, the Deribit exchange provides the deepest options market, and the DVOL index (Deribit's 30-day constant-maturity IV index, analogous to VIX for equities) has become a widely-referenced benchmark.

Implied volatility is a forward-looking, market-consensus forecast. Its advantages:

Incorporates information from all participants simultaneously
Responds in real-time to news flow
Reflects risk premium as well as pure volatility expectation

Its limitations for crypto volatility prediction:

Options markets are less liquid than spot, particularly at longer maturities and away from the money
IV includes a variance risk premium — the market systematically prices options above realized volatility on average, meaning IV is a biased (upward) forecast of subsequent RV
During low-liquidity periods (weekends, holiday windows), IV can move on thin volume in ways that distort the signal

In practice, combining IV with a statistical model (GARCH or HAR-RV) often outperforms either alone. A simple approach: use IV as a prior and update it with recent realized variance data.

4. Machine Learning Approaches

The volatility model crypto research space has expanded significantly into ML territory over the past few years. Common approaches include:

LSTM networks: Long Short-Term Memory models can capture non-linear dependencies in return sequences. They've shown promise for short-horizon (1–3 day) BTC volatility forecasting, particularly when trained on large feature sets that include on-chain metrics, funding rates, and sentiment data.
Random forests and gradient boosting: Tree-based models handle mixed feature types well and don't require distributional assumptions. They can be useful for medium-horizon (1–4 week) forecasts where macro variables and on-chain signals carry more weight.
Transformer-based models: Attention mechanisms can capture long-range dependencies in the return series. Early results are promising but require careful regularization to avoid overfitting on the relatively short history of crypto data.

The honest assessment: ML models for volatility forecasting are harder to validate, more prone to overfitting, and less interpretable than parametric models. They tend to perform best when the feature set is rich and carefully constructed, and worst when trained naively on price data alone. For most practitioners, a well-specified GARCH or HAR-RV model will outperform a poorly-specified LSTM on out-of-sample data.

Combining Forecasts

The empirical literature on volatility forecasting consistently finds that forecast combinations outperform individual models. A simple equally-weighted average of GARCH, HAR-RV, and IV forecasts typically beats any single component on out-of-sample QLIKE or MSE metrics.

More sophisticated combination methods include:

Inverse MSE weighting: Weight each model by the inverse of its recent mean squared error, updated on a rolling basis
Regime-conditional weighting: Use different model weights depending on the current volatility regime — GARCH may be more accurate in calm regimes while HAR-RV captures expansion phases better
Bayesian model averaging: Assign posterior weights based on model likelihood given recent data

The regime-conditional approach is particularly relevant for Bitcoin, given how dramatically its statistical properties shift between accumulation and expansion phases.

How Volatility Forecasts Feed Into Regime Classification

This is where bitcoin volatility forecasting becomes directly actionable rather than academically interesting. Volatility forecasts don't just tell you how much BTC might move — they're a primary input into identifying which market regime you're in.

The connection works in both directions:

Volatility → Regime: A rising volatility forecast, particularly when it crosses specific thresholds, is a leading indicator of regime transition. A calm market where 30-day realized volatility is running at 35% annualized behaves fundamentally differently from one running at 80%. Position sizing, entry timing, and risk limits should all be calibrated to the current regime, not a static historical average.

Regime → Volatility model selection: As noted above, different models perform differently across regimes. A regime-aware forecasting system dynamically adjusts model weights or specifications based on the detected state. This is the approach RegimeRisk takes — integrating volatility signal with broader regime classification to produce context-aware risk assessments rather than point-in-time vol numbers.

The current market environment illustrates this well. BTC's move to $81,500 on tokenization momentum and macro tailwinds represents a potential regime transition from the range-bound conditions that characterized early 2026. Whether that move signals the beginning of a sustained expansion phase or a false breakout is precisely the question that regime classification — informed by volatility forecasts — is designed to answer. Understanding how regime transitions are identified in practice provides important context for interpreting these signals.

For traders, the practical implication is that a single volatility number isn't enough. You need to know whether current volatility is elevated relative to the current regime baseline, whether it's trending up or down, and whether the regime itself appears stable or transitioning. These are distinct questions that require distinct analytical tools.

This also connects directly to how macro conditions interact with on-chain and technical signals. As we've explored in our analysis of on-chain metrics and regime detection, volatility forecasts gain predictive power when combined with structural market signals rather than used in isolation.

Evaluation: How to Know If Your Model Is Working

Volatility forecasts are notoriously difficult to evaluate because volatility itself is latent — you never observe it directly, only through realized returns. Standard evaluation metrics:

QLIKE loss: Asymmetric loss function that penalizes underestimation more than overestimation. Preferred in the academic literature for vol model comparison.
MSE on realized variance: Simple and interpretable, but can be dominated by large outliers.
Mincer-Zarnowitz regressions: Regress realized variance on the forecast. Efficiency requires a coefficient of 1 on the forecast and a zero intercept. Deviations indicate systematic bias.
Value-at-Risk backtests: If your vol forecast is used to construct VaR estimates, Kupiec and Christoffersen tests check whether exceedances are correctly calibrated.

For practical trading applications, the most relevant test is often the simplest: does using the volatility forecast improve risk-adjusted returns on your specific strategy, compared to a static vol assumption? Out-of-sample testing on a held-out period is essential.

Key Takeaways

Bitcoin volatility forecasting is a genuinely hard problem, but it's tractable with the right tools. GARCH variants — particularly GJR-GARCH with fat-tailed error distributions — remain competitive baselines for daily horizon forecasting, while HAR-RV models that exploit intraday data often outperform at short horizons. Implied volatility from options markets provides a real-time, market-consensus signal but carries an upward bias due to the variance risk premium, making it most useful when combined with statistical models rather than used alone.

Forecast combination consistently outperforms individual models in out-of-sample testing, and regime-conditional weighting — adjusting model weights based on the current market state — is particularly well-suited to Bitcoin's tendency to shift between structurally distinct phases. The current environment, with BTC recovering toward multi-month highs on macro and tokenization catalysts, is exactly the kind of transition period where static volatility assumptions break down and dynamic, regime-aware forecasting adds the most value.

Finally, volatility forecasts are most useful not as standalone numbers but as inputs to a broader analytical framework. Knowing that 30-day realized vol is trending up matters less than knowing whether that trend is consistent with a regime expansion, a false breakout, or the early stages of a structural shift — distinctions that require integrating volatility signals with the full landscape of market conditions.