Quant Analytics: from scratch to pro · Factors, volatility, regimes and lead-lag · Advanced
Volatility models: EWMA and GARCH
Volatility is not constant: calm and turbulent spells cluster. EWMA and GARCH update a variance estimate each day from the latest return, and GARCH adds a long-run level that forecasts revert to. Persistence sets how long a shock lingers.
In this lesson
- Update an EWMA variance estimate from a new return
- Update a GARCH(1,1) variance and compute its long-run variance
- Compute persistence and the half-life of a volatility shock
- Forecast volatility several days ahead and explain why the forecast reverts
Read first: Volatility: from daily moves to an annual figure, Time series basics: random walks, autocorrelation and stationarity
Daily returns are not equally risky. Quiet weeks are followed by quiet weeks, and turbulent days by more turbulent days. This volatility clustering is one of the most robust facts about market returns (see return distributions and Cont, 2001). A constant-volatility assumption ignores it. EWMA and GARCH model it.
EWMA
The exponentially weighted moving average estimates today's variance as a weighted blend of yesterday's estimate and yesterday's squared return:
var_t = lambda x var_(t-1) + (1 - lambda) x r_(t-1)^2
The decay factor lambda sits between 0 and 1. The RiskMetrics system popularised lambda = 0.94 for daily data (J.P. Morgan and Reuters, 1996). Recent returns weigh most and old ones fade smoothly, with no sharp window edge. See EWMA.
Worked example: EWMA
Yesterday's daily variance estimate is 0.0001 (a volatility of 1%) and the return was -2%. The new variance is 0.94 x 0.0001 + 0.06 x 0.0004, which is 0.000118. Its square root is 1.086% a day, and 17.24% annualised (x sqrt(252)).
The weight on data k days old is (1 - lambda) x lambda^k. A weight halves every ln(0.5) / ln(0.94), which is 11.2 days, and the average age of the data is 1 / (1 - lambda), 16.7 days (counting yesterday's return as one day old). The annualised figure assumes independent daily returns.
GARCH(1,1)
Engle (1982) introduced ARCH, in which the variance depends on past squared shocks, and applied it to United Kingdom inflation from 1958 to 1977. Bollerslev (1986) generalised it by adding past variance. The GARCH(1,1) model is:
var_t = omega + alpha x r_(t-1)^2 + beta x var_(t-1)
alphais the reaction to the latest shock.betais the memory of past variance.omegasets the long-run level:V = omega / (1 - alpha - beta).
EWMA is the special case omega = 0, alpha = 1 - lambda, beta = lambda. GARCH adds a level the variance returns to. See GARCH.
Worked example: GARCH
Take alpha = 0.08, beta = 0.90 and a long-run daily variance of 0.0001 (1% a day, or 15.87% a year). Then omega = 0.0001 x (1 - 0.98), which is 0.000002.
After a return of -3% with yesterday's volatility at 2%, the next-day variance is 0.000002 + 0.08 x 0.0009 + 0.90 x 0.0004, which is 0.000434, a volatility of 2.083%.
Persistence and half-life
The sum alpha + beta is the persistence: 0.98 here. A shock to variance decays by that factor each day, so its half-life is ln(0.5) / ln(0.98), which is 34.3 days. See half-life.
The forecast k days ahead is var = V + (alpha + beta)^k x (var_now - V). From a 2% daily volatility, the forecast ten days out is 1.858%. The chart shows the decay for two persistence values against EWMA, which stays flat. Persistence near 1 means shocks last for months; a sum at or above 1 means no long-run level exists.
Common mistakes
- Fitting GARCH on a short window and trusting the parameters. They are noisy, and persistence close to 1 is a warning sign.
- Reading forecasts as certain. They are conditional averages; the next day can differ widely.
- Forgetting that returns are fat-tailed even after scaling by the model's volatility.
- Using EWMA for long forecasts. It has no mean reversion.
- Mixing units: a daily variance with an annual one.
In practice
Use EWMA as a simple, hard-to-break baseline and GARCH when the long-run level and multi-day forecast matter. Check that fitted values behave: standardised returns (r / sigma) should show little remaining clustering. These estimates feed volatility targeting and Value at Risk. Analytics on FinUties applies measures like these to the markets it covers, so they are at hand when you read a chart.
Volatility forecast after a spike, by model
- GARCH, persistence 0.98
- GARCH, persistence 0.90
- EWMA (flat)
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| new EWMA variance | 0.94 * 0.0001 + 0.06 * 0.02**2 | 0.000118 |
| daily volatility from it | sqrt(ewma_var) | 1.086% |
| annualised | ewma_vol * sqrt(252) | 17.24% |
| half-life of an EWMA weight, in days | ln(0.5) / ln(0.94) | 11.2 |
| average age of data, in days | 1 / (1 - 0.94) | 16.7 |
| omega for long-run daily variance 0.0001 | 0.0001 * (1 - 0.08 - 0.90) | 0.000002 |
| next-day variance after a -3% return | omega + 0.08 * 0.03**2 + 0.90 * 0.02**2 | 0.000434 |
| next-day volatility | sqrt(g_var) | 2.083% |
| persistence | 0.08 + 0.90 | 0.98 |
| half-life of a variance shock, in days | ln(0.5) / ln(persist) | 34.3 |
| volatility forecast 10 days ahead from 2% | sqrt(0.0001 + persist**10 * (0.02**2 - 0.0001)) | 1.858% |
| long-run annual volatility | sqrt(0.0001 * 252) | 15.87% |
Check yourself
EWMA with lambda 0.94: yesterday's daily variance estimate was 0.0001 and today's return is -2%. What is the new variance?
0.94 x 0.0001 + 0.06 x 0.0004 = 0.000118, a volatility of 1.086%.
A GARCH(1,1) has alpha 0.08 and beta 0.90. What is its persistence and the half-life of a variance shock?
Persistence 0.98. Half-life = ln(0.5) / ln(0.98) = 34.3 days.
What long-run variance does omega = 0.000002 imply with alpha 0.08 and beta 0.90?
omega / (1 - alpha - beta) = 0.000002 / 0.02 = 0.0001, a daily volatility of 1%.
How does EWMA differ from GARCH in a 30-day forecast?
EWMA has no long-run level (it is GARCH with omega = 0), so its forecast stays flat at today's value. GARCH reverts towards the long-run variance.
What happens if alpha + beta is 1.01?
The process is not stationary: shocks never fade and there is no finite long-run variance. A fitted value that high on a short sample usually signals a model problem.
Key terms
Exponentially weighted moving averageGARCHHalf-lifeVolatility
Checked facts
Every figure below carries how it was checked: Primary checked (read in the cited primary, official, academic or data source), Two sources agree, or Approximate (an order of magnitude only).
- Two sources agree
Engle (1982) introduced ARCH models and applied them to United Kingdom inflation from 1958 to 1977.
Sources
- Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation. Econometrica 50(4), 987-1007 (Robert F. Engle), 1982. Accessed 9 Nov 2026
- Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31(3), 307-327 (Tim Bollerslev), 1986. Accessed 9 Nov 2026
- RiskMetrics Technical Document (fourth edition). J.P. Morgan and Reuters, New York, 17 December 1996, 1996. Accessed 9 Nov 2026
- Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance 1(2), 223-236 (Rama Cont), 2001. Accessed 9 Nov 2026
- Modelling UK inflation uncertainty, 1958-2006. University of Warwick, Department of Economics (Gianna Boero, Jeremy Smith, Kenneth F. Wallis). Accessed 9 Nov 2026