GARCH Volatility Modeling: Forecasting Time-Varying Market Risk

Key Takeaways

• GARCH models conditional variance as a weighted blend of a long-run floor, yesterday's squared shock, and yesterday's variance.
• In a typical SPY fit, alpha + beta near 0.98 gives a volatility shock a half-life of roughly 35 days.
• Applying plain symmetric GARCH to equities underestimates post-selloff risk because it ignores the leverage effect.
• Portfolio risk systems use GARCH forecasts to size positions and set dynamic VaR limits across regimes.

What It Is

GARCH stands for Generalized Autoregressive Conditional Heteroskedasticity. The model was introduced by Tim Bollerslev in 1986 as a generalization of Robert Engle's 1982 ARCH model. Engle later shared the 2003 Nobel Memorial Prize in Economic Sciences for this work.

The core idea is that the conditional variance of returns is not constant. Instead, it depends on past squared shocks and past variances. That single idea allows GARCH to reproduce the fat tails, persistence, and clustering that real financial returns exhibit and that the constant-variance assumption of Black-Scholes cannot.

The Intuition

If you plot daily returns for any liquid market, you will see long stretches of small moves punctuated by bursts of large moves. The 2008 crisis, the March 2020 Covid drawdown, and the August 2015 yuan devaluation each produced several weeks of elevated variance before settling back down.

A constant-variance model assumes every day draws from the same distribution. That forecast is clearly wrong after a 5 percent down day. GARCH fixes the problem by letting today's variance depend on yesterday's squared return and yesterday's variance, so a large shock raises the expected variance for tomorrow and the effect decays gradually.

How It Works

The GARCH(1,1) model, by far the most widely used specification, has two equations. The mean equation describes the return, and the variance equation describes how conditional variance evolves.

r_t = mu + eps_t
eps_t = sigma_t * z_t,   z_t ~ N(0,1) or Student-t
sigma_t^2 = omega + alpha * eps_(t-1)^2 + beta * sigma_(t-1)^2

Where:

• r_t is the return at time t.
• eps_t is the shock (innovation) at time t.
• sigma_t^2 is the conditional variance.
• omega is a positive constant (the long-run variance floor).
• alpha weights the previous squared shock (the ARCH term).
• beta weights the previous conditional variance (the GARCH term).

The parameters must satisfy omega > 0, alpha >= 0, beta >= 0, and alpha + beta < 1 for the variance to be stationary. The long-run (unconditional) variance is omega / (1 - alpha - beta).

Extensions that matter

Standard GARCH is symmetric: a +2 percent shock and a -2 percent shock move variance by the same amount. Equity markets violate that. The leverage effect means negative returns raise volatility more than positive returns of equal size. Two common fixes:

• GJR-GARCH (Glosten, Jagannathan, Runkle) adds an indicator term that activates only when the past shock is negative, so sigma_t^2 = omega + (alpha + gamma * I_neg) * eps_(t-1)^2 + beta * sigma_(t-1)^2.
• EGARCH (Nelson) models log-variance, which guarantees positivity and allows asymmetric response without non-negativity constraints on parameters.

IGARCH (integrated GARCH) forces alpha + beta = 1, giving a unit-root in variance and infinite unconditional variance, sometimes useful for long-memory series.

Worked Example

Suppose you fit GARCH(1,1) on SPY daily returns and estimate omega = 1e-6, alpha = 0.09, beta = 0.89.

The long-run daily variance is 1e-6 / (1 - 0.09 - 0.89) = 1e-4, or a long-run daily standard deviation of 1.0 percent. That implies roughly 16 percent annualized, which lines up with typical SPY realized volatility.

Now suppose yesterday's return shock was -3% (so eps_(t-1)^2 = 0.0009) and yesterday's variance was sigma_(t-1)^2 = 2e-4 (standard deviation 1.41 percent, an elevated day). Today's variance forecast:

sigma_t^2 = 1e-6 + 0.09 * 0.0009 + 0.89 * 2e-4
          = 1e-6 + 8.1e-5 + 1.78e-4
          = 2.60e-4

Today's forecast standard deviation is about 1.61 percent, higher than yesterday. A fresh shock amplifies expected volatility, and the effect decays at rate alpha + beta = 0.98 per day, giving a half-life of roughly 35 days.

Common Mistakes

1. Treating GARCH as a return forecast. GARCH forecasts the variance of returns, not their direction. Using fitted r_t values to trade direction is a category error. The signal is risk, not alpha.

2. Ignoring the leverage effect in equities. Plain symmetric GARCH systematically underestimates risk after selloffs in stock indices. If you model SPY or single-name equities without at least GJR or EGARCH, your VaR numbers will lag reality precisely when they matter most.

3. Applying GARCH to non-stationary series. GARCH assumes the return series is stationary and its mean is constant or slowly varying. Fitting GARCH to raw prices, or to a series with structural breaks (a currency peg that broke, a merger), will produce unstable estimates. Always difference or log-difference first.

4. Assuming Gaussian innovations when tails are fat. Real return shocks have excess kurtosis even after GARCH has removed clustering. Using Student-t or skewed-t innovations produces more realistic tail forecasts and is the default in most professional implementations.

5. Overfitting with high-order models. GARCH(1,1) is the workhorse for a reason. GARCH(2,2) or higher rarely beats (1,1) out of sample and often produces unstable parameters. Start simple, add asymmetry before adding lags, and validate with rolling out-of-sample tests.

Frequently Asked Questions

Q: What is GARCH volatility modeling in simple terms? GARCH is a statistical model that says today's market variance depends on yesterday's variance and yesterday's squared return, so volatility has memory instead of resetting each day to a constant level.

Q: How does GARCH volatility modeling affect investment decisions? It lets risk managers set dynamic position sizes, VaR limits, and options hedges that respond to current volatility conditions rather than using a single long-run average that lags badly during crises.

Q: What is a real-world example of GARCH volatility modeling? After the March 2020 Covid crash produced days of minus 9 percent returns, a GARCH(1,1) fit to SPY would have raised the next-day volatility forecast from roughly 1 percent to over 4 percent within a week, prompting traders to cut leverage accordingly.

Q: How can investors use GARCH volatility modeling? By fitting GARCH to historical returns and using the one-step-ahead variance forecast to scale position sizes inversely with current risk, so exposure drops automatically when volatility spikes and rebuilds as it decays.

Q: How is GARCH volatility modeling different from historical volatility? Historical volatility gives equal weight to every return in the lookback window and then drops observations abruptly. GARCH weights recent shocks more heavily and lets their influence decay exponentially, producing smoother and more forward-looking variance estimates.

Sources

1. Engle, R.F. (1982). "Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation." Econometrica 50(4), 987-1007. http://www.econ.uiuc.edu/~econ536/Papers/engle82.pdf
2. Bollerslev, T. (1986). "Generalized Autoregressive Conditional Heteroskedasticity." Journal of Econometrics 31, 307-327. https://public.econ.duke.edu/~boller/Published_Papers/joe_86.pdf
3. V-Lab (NYU Stern). "GARCH Volatility Documentation." https://vlab.stern.nyu.edu/docs/volatility/GARCH
4. V-Lab (NYU Stern). "GJR-GARCH Volatility Documentation." https://vlab.stern.nyu.edu/docs/volatility/GJR-GARCH

Disclaimer

This article is educational content only and is not financial advice. Nothing here is a recommendation to buy, sell, or hold any security. Consult a licensed advisor before making investment decisions.