GARCH, short for generalised autoregressive conditional heteroskedasticity, is a family of models for forecasting volatility that changes over time. It captures volatility clustering: large moves tend to be followed by large moves, and calm periods by calm ones.
The GARCH(1,1) model
σ2t = ω + α x ε2t-1 + β x σ2t-1, where today's variance depends on a long-run base (ω), yesterday's squared surprise return (α) and yesterday's variance (β). It was introduced by Bollerslev (1986), building on Engle's ARCH model (1982).
An illustrative one-day update
| Input | Value |
|---|---|
| ω, α, β | 0.000002, 0.08, 0.90 |
| Yesterday's volatility | 1.0%, so variance 0.0001 |
| Yesterday's return | -3%, so squared 0.0009 |
| Today's variance | 0.000002 + 0.08 x 0.0009 + 0.90 x 0.0001 = 0.000164 |
| Today's volatility | about 1.28% |
One large move raised the forecast volatility from 1.0% to 1.28%. Without further shocks it decays back toward the long-run level of √(ω / (1 - α - β)) = 1.0%. The numbers are made up.
Uses
- Volatility forecasts for risk measures such as VaR and for volatility targeting.
- Time-varying inputs for a covariance matrix.
- Comparison with implied volatility from option prices.
Limits
- Basic GARCH treats a rise and a fall of equal size alike, while real markets show larger volatility after falls. Variants such as EGARCH and GJR-GARCH allow this.
- Parameters can shift between regimes. See hidden Markov models.
- It forecasts the size of moves, not their direction.
Frequently asked questions
What is a GARCH model?
It is a statistical model that forecasts time-varying volatility by combining a long-run level, recent shocks and the previous variance.
What is volatility clustering?
The tendency of large price changes to be followed by large changes, and small changes by small ones.
What is the GARCH(1,1) formula?
Variance today = a constant + alpha x yesterday's squared return shock + beta x yesterday's variance.
What are GARCH models used for?
Forecasting volatility for risk management, volatility targeting, option pricing and covariance estimation.