Volatility clustering and the GARCH model
Daily stock returns are almost uncorrelated, yet calm and turbulent periods cluster. The GARCH model describes this with one recursion for the variance; a simulation shows its consequences for value at risk and for the backtesting of bank risk models.
Anyone who held shares in March 2020 saw the pattern at first hand. After weeks of small daily movements, stock indices moved by several per cent a day, up and down, for about a month, and then gradually calmed. The VIX index, which measures the volatility expected in the US stock market, reached its highest closing level on record, about 83, on 16 March 2020. The direction of each day's move remained as unpredictable as before. Its size did not: large moves followed large moves.
Benoit Mandelbrot described the same pattern in cotton prices in 1963: large changes tend to be followed by large changes, of either sign, and small changes by small changes. It is a form of dependence that correlation does not detect, as discussed in correlation, independence and transfer entropy: today's return says almost nothing about the sign of tomorrow's, but a great deal about its size.
The pattern matters well beyond trading floors. A pension fund that reports the risk of its portfolio, a bank that holds capital against market losses, and a saver who wants to know how far a fund could fall all depend on an estimate of volatility. If volatility changes over time, an estimate from a calm year understates the risk at the start of a turbulent one. This article presents the standard model of the effect and examines its consequences for value at risk.
The relevant literature is as follows.
- Mandelbrot (1963) documented heavy tails and volatility clustering in commodity prices.
- Engle (1982) introduced the autoregressive conditional heteroscedasticity (ARCH) model, applied first to the variance of UK inflation. Engle shared the 2003 Nobel memorial prize in economics with Clive Granger, in his case for methods of analysing time series with time-varying volatility.
- Bollerslev (1986) generalised it to GARCH, which adds the previous variance to the recursion, and Bollerslev (1987) combined it with Student's errors for heavier tails.
- J.P. Morgan and Reuters (1996) published the RiskMetrics methodology, whose exponentially weighted variance with a decay factor of 0.94 for daily data became an industry standard.
- Cont (2001) summarised the "stylised facts" of asset returns: near-zero autocorrelation of returns, slowly decaying autocorrelation of absolute and squared returns, and heavy tails.
- The Basel Committee (1996) set out the backtesting of banks' value-at-risk models, and the Basel Committee (2019), in its revised market-risk framework, replaced the 99% value at risk with a 97.5% expected shortfall for internal models.
The model
Write the daily return as , where is an independent standard normal shock and the volatility for day , known at the end of day . The GARCH(1,1) model sets
Each day's variance is a weighted combination of three things: a long-run level, yesterday's squared surprise and yesterday's variance. A large move yesterday raises today's variance through ; a turbulent period persists through . If , the variance returns towards its long-run value .
The sum is the persistence of volatility. A shock to the variance decays like after days, so its half-life is . With and the persistence is 0.98 and the half-life about 34 trading days, roughly a month and a half. Estimates for daily stock-index returns often put the persistence between about 0.97 and 0.99, which matches the experience of March 2020: the turbulence faded over weeks, not days.
The RiskMetrics method is the special case , and . Its persistence is exactly 1, so it has no long-run level to return to: a convenient approximation for the next few days, less suitable for horizons of months.
Heavy tails from normal shocks
Even with normal shocks , the returns of a GARCH process have heavier tails than the normal. Calm periods with small and turbulent periods with large are mixed together, and a mixture of normals with different variances has more mass in the centre and in the tails than a single normal. For GARCH(1,1) the kurtosis is
provided the denominator is positive. With the parameters above it is about 4.4, against 3 for the normal. Observed daily returns usually show more kurtosis than a normal GARCH produces, which is why heavier-tailed shocks, such as Student's or the exponential power family discussed in the γ-order generalized normal distribution, are often used for . Part of the heavy tails comes from changing volatility, and part from the shocks themselves.
Value at risk and its backtesting
A bank's one-day 99% value at risk (VaR) is the loss that should be exceeded on only one day in a hundred. Supervisors check it by backtesting: counting the days on which the actual loss exceeded the reported VaR. Over 250 trading days about 2.5 exceptions are expected. The Basel framework places a model in a green zone with up to four exceptions, a yellow zone with five to nine and a red zone with ten or more, and a model in the yellow or red zone attracts a higher capital requirement.
Volatility clustering has a direct effect on this count. A VaR based on the standard deviation of the past year reacts slowly. After a turbulent period it remains too high for months and records almost no exceptions; at the start of a turbulent period it is too low and records many in quick succession. In March 2020 many banks recorded more exceptions in a few weeks than in the preceding years, and supervisors in Europe and the United States granted temporary relief from the capital consequences. A VaR built on a GARCH volatility rises after the first large move and spreads its exceptions more evenly over time.
The revised framework of 2019 moved internal models from the 99% VaR to the 97.5% expected shortfall, the average loss on the worst 2.5% of days, partly because VaR says nothing about how large the losses beyond it are.
The figure below simulates nine years of daily returns and backtests both methods. Returns are almost uncorrelated while their squares are not; volatility clusters; and a VaR based on the past year's standard deviation produces exceptions in bunches, often reaching the red zone, while a VaR based on the GARCH volatility does not.
Suggested settings:
- Clustering. With the defaults (, ), the return series alternates between calm and turbulent stretches, and the conditional band widens and narrows with them. The lag-1 autocorrelation of returns is close to zero, while that of squared returns is about 0.2 and decays slowly over the following lags.
- Constant volatility. Turn on "Constant volatility". The overall variance is the same, but the clusters disappear, the autocorrelation of squares falls to about zero and the kurtosis returns to about 3.
- Backtesting. Return to GARCH data and press "New 9 years of data" several times. Over 2,000 days, about 20 exceptions are expected. The rolling-year VaR records about 30 on average, and in more than half of the runs its worst 250-day window reaches ten or more, the red zone. The GARCH VaR records about 20, and its worst window rarely reaches the red zone.
- Long and short memory. Set and : the half-life grows to more than 100 days and turbulent periods last for months. Set and : shocks are larger but fade within a few days.
The long-run variance and the decay of shocks. Take expectations of the recursion. Because , a stationary variance satisfies , so . For the forecast days ahead, , since . Iterating gives : the excess variance decays geometrically at the rate of the persistence, which gives the half-life above.
Volatility clustering is one of the most robust regularities in financial data, and GARCH describes it with three parameters. The practical consequences are concrete:
- A risk estimate based on a fixed window lags behind changes in volatility, and its errors arrive in clusters.
- The persistence indicates how long a turbulent period is likely to last, which matters for funding plans and margin calls as well as for capital.
- Changing volatility explains part of the heavy tails of returns; the remainder calls for a heavier-tailed shock distribution.
References
- B. Mandelbrot, "The variation of certain speculative prices", Journal of Business, 36(4), 394-419, 1963.
- R. F. Engle, "Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation", Econometrica, 50(4), 987-1007, 1982.
- T. Bollerslev, "Generalized autoregressive conditional heteroskedasticity", Journal of Econometrics, 31(3), 307-327, 1986.
- T. Bollerslev, "A conditionally heteroskedastic time series model for speculative prices and rates of return", Review of Economics and Statistics, 69(3), 542-547, 1987.
- J.P. Morgan and Reuters, RiskMetrics Technical Document, 4th edition, 1996.
- R. Cont, "Empirical properties of asset returns: stylized facts and statistical issues", Quantitative Finance, 1(2), 223-236, 2001.
- Basel Committee on Banking Supervision, Supervisory framework for the use of "backtesting" in conjunction with the internal models approach to market risk capital requirements, 1996.
- Basel Committee on Banking Supervision, Minimum capital requirements for market risk, 2019.