GARCH Model

Quantitative Finance · advanced · CC-BY-4.0

The GARCH (Generalized Autoregressive Conditional Heteroskedasticity) model, introduced by Tim Bollerslev in 1986, is a statistical time series model that captures the well-documented tendency of financial asset return volatility to cluster—periods of high volatility tend to be followed by more high volatility—by modeling conditional variance as a function of past squared residuals and past conditional variances. It extends Engle's ARCH model and is the foundational framework for volatility modeling in quantitative finance.

Key takeaways

The GARCH(1,1) model—the most widely used specification—models the conditional variance as: σ²_t = ω + α × ε²_(t-1) + β × σ²_(t-1), where ω is the long-run variance intercept, α captures the ARCH effect (impact of past shocks), and β captures the GARCH effect (persistence of past volatility); the sum α + β measures volatility persistence.
Volatility clustering—the empirical observation that large price moves are more likely to be followed by large moves, and small moves by small moves—is the key stylized fact that GARCH models are designed to capture; this clustering violates the constant volatility assumption of Black-Scholes.
When α + β is close to 1 (high persistence), volatility mean-reverts very slowly to its long-run average ω/(1−α−β); an α + β of 0.99 implies an extremely slow mean reversion that is characteristic of equity market volatility, where volatility shocks can persist for months.
Extensions of the basic GARCH model include EGARCH (exponential GARCH, capturing asymmetric responses where negative returns increase volatility more than positive returns—the 'leverage effect'), TARCH/GJR-GARCH (threshold models), IGARCH (integrated, where α+β=1), and multivariate GARCH models (DCC-GARCH) for joint volatility modeling.
GARCH models are used for value-at-risk (VaR) calculation, options pricing under stochastic volatility, portfolio risk management, and forecasting future volatility; they outperform constant volatility models in forecasting one-step-ahead conditional variance but have limited accuracy for long-horizon volatility forecasting.

Explanation

The GARCH model represents one of the most significant contributions to empirical finance and quantitative risk management of the late 20th century. Its development by Tim Bollerslev (1986) built on Robert Engle's ARCH (AutoRegressive Conditional Heteroskedasticity) model (1982), for which Engle received the 2003 Nobel Prize in Economics. The core insight is simple but profound: financial return volatility is not constant over time—it exhibits serial dependence (clustering) that can be modeled and, to a degree, forecasted. Incorporating this time-varying volatility structure dramatically improves the statistical realism of financial models relative to the constant-variance assumptions of Black-Scholes and traditional portfolio theory.

The canonical GARCH(1,1) specification models the conditional variance of returns as: σ²_t = ω + α × ε²_(t-1) + β × σ²_(t-1), where σ²_t is today's conditional variance, ε_(t-1) is the previous period's standardized return shock (innovation), and σ²_(t-1) is the previous period's conditional variance. The parameter ω (omega) is a positive constant that anchors the conditional variance to a long-run level; α (alpha) is the ARCH parameter measuring how quickly current volatility responds to recent shocks; β (beta) is the GARCH parameter measuring the persistence of conditional variance. The unconditional long-run variance is ω/(1−α−β), provided α + β < 1 (the stationarity condition).

In practice, GARCH models estimated on equity market data typically show high persistence: α + β values of 0.97–0.99 are common, implying that volatility shocks decay very slowly. For daily S&P 500 returns, a typical GARCH(1,1) estimate might yield ω ≈ 0.000002, α ≈ 0.09, β ≈ 0.90, giving α + β = 0.99 and a long-run daily variance of 0.000002 / (1 − 0.99) = 0.0002, corresponding to a daily volatility of about 1.41% or annualized volatility of 22.4%. The high β implies that shocks to volatility persist for a very long time—after a volatility spike, the model predicts elevated volatility for months rather than reverting to the long-run mean within days.

A critical limitation of the symmetric GARCH model is its failure to capture the leverage effect—the empirical observation that equity market volatility responds more strongly to negative returns than to positive returns of equal magnitude. This asymmetry arises because declining stock prices increase financial and operating leverage (as the equity cushion shrinks relative to debt), increasing the riskiness of the firm. EGARCH (Exponential GARCH), developed by Nelson (1991), addresses this by modeling log conditional variance: ln(σ²_t) = ω + α × (|ε_(t-1)/σ_(t-1)| − E[|ε/σ|]) + γ × (ε_(t-1)/σ_(t-1)) + β × ln(σ²_(t-1)), where the γ term captures asymmetry. The GJR-GARCH model, alternatively, uses an indicator function to add extra volatility impact for negative shocks. Both models consistently fit equity return data better than the symmetric GARCH.

Multivariate extensions of GARCH, particularly the DCC-GARCH (Dynamic Conditional Correlation GARCH) model introduced by Engle (2002), are widely used in portfolio risk management. DCC-GARCH allows the correlation between asset returns to vary over time alongside their individual volatilities, capturing the well-documented phenomenon of correlation spikes during market stress. This is critical for realistic VaR estimation and stress testing, as assuming constant correlations (as in traditional Markowitz mean-variance optimization) dramatically underestimates tail risk in multi-asset portfolios during crises, when correlations among risky assets typically converge toward 1.

Formula

GARCH(1,1): σ²_t = ω + α × ε²_(t-1) + β × σ²_(t-1); Long-run variance = ω / (1 − α − β), where α + β < 1

Example

A risk manager at a hedge fund estimates a GARCH(1,1) model on daily S&P 500 returns from 2010–2024, obtaining parameters: ω = 0.0000015, α = 0.08, β = 0.91 (persistence = 0.99, long-run daily variance = 0.00015, long-run annualized vol = 19.4%). On March 16, 2020—during the COVID-19 market crash—the S&P 500 fell 12% (the largest single-day decline in the sample). The squared shock ε²_(t) = (0.12)² = 0.0144. The model updates the conditional variance: σ²_(t+1) = 0.0000015 + 0.08 × 0.0144 + 0.91 × σ²_t. If σ²_t was already elevated at 0.001 (daily vol of 3.16%) before the crash, the post-crash conditional variance jumps to 0.0000015 + 0.001152 + 0.000910 = 0.0021, corresponding to daily volatility of 4.58% (annualized: 72.7%). This accurately captures the volatility spike observed in the VIX during March 2020 (which peaked above 80). The high persistence (β = 0.91) means that this elevated volatility estimate persists in the model for many weeks, consistent with actual market behavior where VIX remained elevated well into Q2 2020.

GARCH Model

Key takeaways

Explanation

Formula

Example

Related terms