GARCH Models: Forecasting Volatility in Practice

Key Takeaway

GARCH(1,1) remains the workhorse of volatility forecasting four decades after its introduction, capturing over 90% of conditional variance dynamics with just three parameters. While asymmetric extensions like EGARCH and GJR-GARCH improve performance during market stress by modeling the leverage effect, Hansen and Lunde (2005) found that no model in a comparison of 330 GARCH variants consistently outperformed a well-estimated GARCH(1,1) for daily exchange rate volatility. The practical lesson is clear: model parsimony often beats model complexity in out-of-sample forecasting.

From Constant to Conditional Volatility

Before Robert Engle published his seminal 1982 paper, financial econometrics treated volatility as constant. Portfolio optimization used a single variance estimate derived from the full sample, risk measures assumed stable distributions, and options were priced under the assumption that volatility was a known, fixed parameter. Anyone who had watched markets for more than a few months knew this was wrong. Volatility clusters: calm periods persist, and turbulent periods persist. The crash of October 1987, the 1997 Asian financial crisis, and the 2008 global financial crisis all exhibited dramatic volatility clustering that constant-variance models could not capture.

Engle (1982) formalized this observation with the Autoregressive Conditional Heteroskedasticity (ARCH) model. Rather than treating variance as fixed, the ARCH model allows the conditional variance at time t to depend on past squared returns. In the simplest ARCH(1) specification:

h(t) = omega + alpha * epsilon(t-1)^2

where h(t) is the conditional variance, omega is a baseline variance level, epsilon(t-1) is the previous period's return shock, and alpha governs how strongly yesterday's shock affects today's variance estimate. When alpha is large, a big return shock (positive or negative) causes a large increase in the next period's estimated variance. When alpha is small, the model is closer to constant variance.

The ARCH model earned Engle a share of the 2003 Nobel Prize in Economics. But the original formulation had a practical limitation: to capture the slow decay of volatility clustering, you needed many lagged squared returns (high-order ARCH), which required estimating many parameters and often produced unstable estimates.

The GARCH(1,1) Breakthrough

Bollerslev (1986) solved this parsimony problem with the Generalized ARCH model, or GARCH. The key insight was to include the lagged conditional variance itself as a predictor:

h(t) = omega + alpha * epsilon(t-1)^2 + beta * h(t-1)

This single equation, GARCH(1,1), captures both the immediate impact of a return shock (through alpha) and the persistence of past volatility (through beta). The parameter beta acts as an exponential smoothing weight on the entire history of squared returns, allowing the model to generate long-persisting volatility clusters with only three parameters.

The sum alpha + beta is the persistence parameter. When this sum is close to 1, shocks to volatility die out slowly, and the unconditional variance omega / (1 - alpha - beta) is large. When the sum equals 1 exactly, the process is integrated GARCH (IGARCH), meaning volatility shocks never fully dissipate. Empirically, estimates for daily equity returns typically yield alpha + beta between 0.97 and 0.995, indicating very high persistence.

The following table shows typical GARCH(1,1) parameter estimates for major asset classes using daily returns:

The half-life column shows how many days it takes for a volatility shock to decay to half its initial impact, calculated as ln(0.5) / ln(alpha + beta). Higher persistence means shocks reverberate longer, which matters for risk management horizons and options pricing.

The Leverage Effect: Why Down Moves Amplify Volatility

One phenomenon that GARCH(1,1) misses is the asymmetry of volatility response to positive versus negative returns. Empirically, negative return shocks increase subsequent volatility more than positive shocks of the same magnitude. This is the leverage effect, first documented by Black (1976), who hypothesized that falling stock prices increase a firm's leverage ratio, making equity more volatile.

Two major extensions address this asymmetry.

Nelson (1991) proposed the Exponential GARCH (EGARCH) model, which models the logarithm of variance rather than variance itself:

ln h(t) = omega + alpha * [|z(t-1)| - E|z(t-1)|] + gamma * z(t-1) + beta * ln h(t-1)

where z(t-1) is the standardized residual. The parameter gamma captures the asymmetry: when gamma is negative, negative shocks increase volatility more than positive shocks. Because the model operates on the log scale, it automatically ensures that the conditional variance is always positive without parameter constraints.

Glosten, Jagannathan, and Runkle (1993) proposed the GJR-GARCH model, which adds an indicator function:

h(t) = omega + (alpha + gamma * I(t-1)) * epsilon(t-1)^2 + beta * h(t-1)

where I(t-1) equals 1 when epsilon(t-1) is negative and 0 otherwise. The parameter gamma captures the additional volatility impact of negative shocks. For the S&P 500, typical estimates give gamma around 0.10 to 0.15, meaning that a negative 2% return increases the next day's conditional variance roughly 50-75% more than a positive 2% return.

The following table compares model fit across these specifications for S&P 500 daily returns (1990-2024):

The asymmetric models (EGARCH, GJR-GARCH, TGARCH) consistently outperform symmetric GARCH on information criteria. Adding a second GARCH lag (GARCH(2,1)) provides almost no improvement, confirming that the leverage effect matters more than additional lags.

Parameter Estimation in Practice

GARCH parameters are typically estimated by maximum likelihood. Under the assumption that standardized residuals follow a normal distribution, the log-likelihood function for a sample of T observations is:

L = -0.5 * sum[ln(h(t)) + epsilon(t)^2 / h(t)]

In practice, financial returns exhibit fatter tails than the normal distribution, so the Student-t distribution or the Generalized Error Distribution (GED) are commonly used instead. The choice of error distribution affects the estimated parameters and the quality of tail risk forecasts.

Several practical considerations matter for robust estimation. The sample size should be at least 1,000 daily observations (roughly four years) to obtain stable parameter estimates. The optimization routine should use analytical gradients where available, and results should be checked against multiple starting values to avoid local optima. Standard errors should be computed using the robust (sandwich) estimator of Bollerslev and Wooldridge (1992), which remains valid even when the error distribution is misspecified.

A common pitfall is interpreting very high beta estimates (above 0.95) without considering structural breaks. If the sample spans a period that includes a fundamental regime change (such as the transition from high to low inflation), the GARCH model will attribute the resulting variance shift to extreme persistence, inflating beta and reducing the model's forecast accuracy.

Forecast Accuracy: What Works in Practice

Hansen and Lunde (2005) conducted the most comprehensive comparison of GARCH-type models, evaluating 330 different specifications for forecasting the daily volatility of IBM stock returns and DM/USD exchange rate returns. Their findings were surprisingly definitive:

For exchange rate data, no model significantly outperformed GARCH(1,1). For equity data, asymmetric models (EGARCH, GJR-GARCH) provided statistically significant improvements over symmetric GARCH. The leverage effect is more pronounced in equity markets, where the negative return-volatility correlation is stronger, than in currency markets.

The following table summarizes out-of-sample forecast accuracy measured by Mean Squared Error (MSE) relative to GARCH(1,1) (normalized to 1.00):

Values below 1.00 indicate better forecast accuracy than GARCH(1,1). The pattern is consistent: asymmetric models improve equity volatility forecasts by 6-7%, offer marginal gains for bonds, and barely matter for currencies. More complex symmetric models rarely help.

Applications in Risk Management and Options Pricing

GARCH models serve two primary practical functions in modern finance.

In risk management, GARCH-based Value at Risk (VaR) and Expected Shortfall (ES) calculations condition the risk estimate on the current volatility regime. During calm periods, a GARCH-based VaR tightens, allowing portfolios to take larger positions within the same risk budget. During turbulent periods, it expands, automatically reducing position sizes. This conditional approach produces more accurate risk forecasts than unconditional methods, particularly at the one-day and 10-day horizons used by regulatory frameworks such as Basel III.

In options pricing, GARCH models bridge the gap between discrete-time econometric modeling and continuous-time option valuation. Duan (1995) developed the locally risk-neutral valuation relationship (LRNVR), which allows GARCH parameters estimated from historical returns to be used for option pricing under the risk-neutral measure. The key insight is that volatility persistence captured by GARCH translates into the term structure of implied volatility: high persistence (alpha + beta near 1) produces a flatter term structure, while lower persistence produces a steeper one. Asymmetric GARCH models additionally generate implied volatility skew, capturing the market's tendency to price downside risk more heavily.

Limitations and Modern Alternatives

GARCH models operate at a single frequency, typically daily. They cannot exploit the information content of intraday data without aggregation, which discards potentially valuable high-frequency signals. Realized volatility measures, constructed from intraday returns, provide more accurate daily variance estimates and can serve as inputs to HAR (Heterogeneous Autoregressive) models that forecast at multiple horizons simultaneously.

GARCH models assume a parametric structure for the conditional variance equation, which may not adapt quickly enough to sudden regime changes such as central bank policy shifts or geopolitical shocks. Regime-switching GARCH models address this by allowing different parameter sets in different market states, at the cost of additional parameters and estimation complexity.

Machine learning approaches, including LSTM neural networks and tree-based models, have shown promise for volatility forecasting by capturing nonlinear patterns that GARCH's linear variance equation misses. However, these models require substantially more data, are prone to overfitting, and lack the interpretability that makes GARCH models useful for regulatory reporting and risk communication.

Despite these limitations, GARCH remains the standard for several reasons: it is computationally fast, theoretically grounded, easy to interpret, and well-supported by decades of empirical evidence. For most practical applications in risk management and derivatives pricing, a well-estimated GARCH(1,1) or GJR-GARCH(1,1) remains the appropriate starting point.

Actionable Takeaway

The GARCH family of models transformed volatility from a fixed parameter into a dynamic, forecastable quantity. For practitioners, the evidence supports a clear hierarchy: start with GARCH(1,1) for its parsimony and robustness, upgrade to GJR-GARCH or EGARCH when modeling equity volatility where the leverage effect matters, and avoid the temptation to add complexity (higher orders, exotic distributions, or regime switching) unless you have strong out-of-sample evidence that the additional parameters improve forecasts. The sum alpha + beta is the single most important diagnostic; values above 0.98 indicate high persistence and suggest that volatility regime shifts will be slow, while values below 0.95 indicate faster mean-reversion and shorter forecast horizons.