Core Path · L6.1.04

GARCH models

14 minGARCHARCHEWMAVolatility forecastingMaximum likelihood

By the end you will be able to

  1. 1Explain how ARCH and GARCH(1,1) turn yesterday's shock and yesterday's variance into today's volatility, and why that reproduces volatility clustering.
  2. 2Fit GARCH(1,1) with normal and Student t shocks by maximum likelihood (M-D08: Maximum likelihood estimation), and read persistence, half-life and the long-run level from the estimates.
  3. 3Show that EWMA is GARCH with ω = 0 and α + β = 1 (integrated GARCH), and what that costs in the likelihood and in forecasts.
  4. 4Forecast volatility from 31 December 2025 over horizons from one day to a year, and compare the term structure with √h scaling.

Why it matters

On 23 June 2016 the model on this page put GBP/USD's volatility for the next day at 13.5% a year. The next fixing, after the referendum result, was −8.16%: a 9.6-sigma day under that forecast. The model could not see the vote coming. What it did next is the point: on the following day it put volatility at 32.5%, and let it decay back over months. Volatility is not a constant to be estimated once. It is a moving quantity to be forecast every evening, and GARCH is the standard way to do it.