Imported from dnkrow/skill (claude-global/volatility-modeling/SKILL.md). Install upstream with npx skills add dnkrow/skill --skill volatility-modeling. Copyright stays with the author.

Volatility Modeling

Volatility — the magnitude of price fluctuations — is arguably the single most important quantity in trading. It drives position sizing, stop placement, option pricing, and regime detection. This skill covers estimation, forecasting, and practical application of volatility in crypto markets.

Why Volatility Matters

Types of Volatility

Historical (Realized) Volatility

Computed from observed past returns. The most common and directly measurable form. Multiple estimators exist with different statistical efficiency.

Implied Volatility

Derived from option prices via Black-Scholes or similar models. Limited in crypto DeFi where liquid options markets are sparse, but available on Deribit for BTC/ETH.

Forecast Volatility

Predicted future volatility from models like EWMA or GARCH. Used for forward-looking position sizing and risk budgets.

Estimation Methods

1. Close-to-Close (Standard Deviation of Log Returns)

The simplest estimator. Compute the standard deviation of log returns and annualize.

import numpy as np

log_returns = np.log(closes[1:] / closes[:-1])
vol_daily = np.std(log_returns, ddof=1)
vol_annual = vol_daily * np.sqrt(365)  # crypto trades 365 days

• Pros: Simple, widely understood.
• Cons: Uses only close prices — ignores intraday range.

2. Parkinson (High-Low Range)

Uses the daily high-low range, which is ~5x more statistically efficient than close-to-close.

hl_ratio = np.log(highs / lows)
vol_parkinson = np.sqrt(np.mean(hl_ratio**2) / (4 * np.log(2))) * np.sqrt(365)

• Pros: More efficient, captures intraday moves.
• Cons: Downward bias with discrete sampling; ignores close-to-close jumps.

3. Garman-Klass (OHLC)

The most efficient single-day OHLC estimator.

hl = np.log(highs / lows)
co = np.log(closes / opens)
gk = np.mean(0.5 * hl**2 - (2 * np.log(2) - 1) * co**2)
vol_gk = np.sqrt(gk) * np.sqrt(365)

• Pros: Best efficiency among OHLC estimators.
• Cons: Assumes no drift; sensitive to opening gaps.

4. Yang-Zhang

Combines overnight (close-to-open) and open-to-close components. Handles gaps properly. Less relevant for 24/7 crypto but useful for tokens with sporadic trading.

5. EWMA (Exponentially Weighted Moving Average)

RiskMetrics approach — no parameters to estimate beyond λ.

lam = 0.94  # RiskMetrics default for daily
ewma_var = np.zeros(len(returns))
ewma_var[0] = returns[0] ** 2
for t in range(1, len(returns)):
    ewma_var[t] = lam * ewma_var[t - 1] + (1 - lam) * returns[t - 1] ** 2
vol_ewma = np.sqrt(ewma_var) * np.sqrt(365)

• λ = 0.94 for daily data (RiskMetrics).
• λ = 0.97 for weekly data.
• Higher λ → smoother, slower reaction to new information.

6. GARCH(1,1)

The workhorse autoregressive volatility model. Captures volatility clustering.

σ²_t = ω + α · r²_{t-1} + β · σ²_{t-1}

• ω: long-run variance weight.
• α: reaction to recent shock (typically 0.05–0.15 for crypto).
• β: persistence (typically 0.80–0.90 for crypto).
• α + β < 1: stationarity constraint.
• Long-run variance: ω / (1 − α − β).

Estimated via maximum likelihood. See references/estimators.md for details.

Volatility Cones

Volatility cones show the percentile distribution of realized volatility at different lookback windows, revealing whether current vol is historically high or low.

Construction

1. Get 1+ years of daily data.
2. For each lookback window (5, 10, 20, 60, 120 days):
   • Compute rolling realized volatility.
   • Extract percentiles: 5th, 25th, 50th, 75th, 95th.
3. Plot percentiles vs window length — the "cone" shape.
4. Overlay current realized vol at each window.

Interpretation

• Current vol > 75th percentile: historically elevated — expect mean reversion.
• Current vol < 25th percentile: historically compressed — expect expansion.
• Cone narrowing at longer windows: vol mean-reverts over longer horizons.

See references/volatility_cones.md for full methodology and worked examples.

Crypto Volatility Characteristics

Crypto vol differs from traditional assets in important ways:

Regime Classification by Volatility

Volatility Forecasting

EWMA Forecast

Simple and effective. The current EWMA variance estimate is the 1-step forecast. Multi-step forecasts are flat (same as 1-step).

GARCH Forecast

GARCH produces a term structure of variance forecasts:

σ²_{t+h} = V_L + (α + β)^h · (σ²_t − V_L)

Where V_L = ω / (1 − α − β) is the long-run variance.

• Short-horizon forecasts reflect current conditions.
• Long-horizon forecasts converge to long-run variance.
• The speed of convergence depends on α + β (persistence).

See scripts/vol_forecast.py for a working implementation.

Practical Applications

Position Sizing with Volatility

# Vol-target position sizing
target_vol = 0.02  # 2% daily portfolio vol target
current_vol = 0.05  # 5% daily asset vol (annualized ~95%)
weight = target_vol / current_vol  # = 0.40 → 40% allocation

See the position-sizing skill for complete integration.

ATR-Based Stop Placement

atr_14 = talib.ATR(highs, lows, closes, timeperiod=14)
stop_distance = 2.0 * atr_14[-1]  # 2x ATR stop
stop_price = entry_price - stop_distance  # for longs

Vol-Regime Strategy Selection

vol_percentile = current_vol_percentile(token, window=30)
if vol_percentile < 25:
    strategy = "mean_reversion"
elif vol_percentile > 75:
    strategy = "momentum_breakout"
else:
    strategy = "balanced"

Files

References

Scripts

Related Skills

• regime-detection — Classify market regimes using volatility as a key input.
• position-sizing — Scale positions inversely with volatility.
• risk-management — Portfolio-level vol targeting and risk budgets.
• pandas-ta — ATR and Bollinger Bands are volatility-based indicators.
• custom-indicators — Build crypto-specific volatility indicators.

Dependencies

uv pip install pandas numpy scipy

Optional for live data:

uv pip install httpx