MerQuR+Statistics Glossary

Statistics Glossary → GARCH (Volatility Model)

GARCH (Volatility Model)analysis

A conditional variance model that describes and forecasts how the volatility (variance) of a return series changes over time, taking into account the clustering of calm and turbulent periods.

Concept

Forecasting the average change in financial returns is very hard, but volatility shows a clear pattern: large swings follow large swings and calm days follow calm days (volatility clustering). Models that assume a constant variance cannot see this; they therefore understate risk in turbulent periods and overstate it in calm ones. GARCH answers the question 'how much risk is there today?' by linking each day's variance to past shocks and past variance.

In the most common form, GARCH(1,1), today's conditional variance has three parts: σ²_t = ω + α·ε²_{t−1} + β·σ²_{t−1}. α shows the effect of yesterday's shock (surprise) on today's risk, and β how much of yesterday's volatility carries over. α + β is the persistence: the closer it is to 1, the longer a shock's effect lasts; the half-life is the time it takes for a shock's effect to halve. EGARCH and GJR-GARCH also model the asymmetric effect whereby bad news (negative shocks) raises volatility more than good news.

The model works with a return series; price levels must first be converted to returns. For stationarity α + β must be below 1; if it is very close to 1, shocks are almost permanent. Financial returns usually have heavier tails than the normal distribution, in which case a Student-t distribution is more suitable. In a good model no autocorrelation remains in the standardised residuals or in their squares.

When to use it

  • To measure risk, forecast volatility or compute value at risk for series whose volatility changes over time, such as stock, currency or commodity returns.
  • If bad news is thought to raise volatility more, use EGARCH or GJR-GARCH; if returns are heavy-tailed, use a Student-t error distribution.
  • If the mean level rather than the variance is to be forecast, use ARIMA; if the dynamic relationship between several series is being studied, VAR is preferred.

An everyday example

A small firm that buys materials from abroad follows the daily percentage changes in the exchange rate. In some weeks the rate moves only 0.1% a day; after a political event, however, 2% jumps appear for days on end and the turbulence lasts a while. A GARCH model captures this clustering: if α + β comes out at, say, 0.95, it takes about 13 to 14 trading days for a shock's effect to halve. Looking at the model's forecast of tomorrow's volatility, the firm can decide whether to make its payment now or wait until things calm down.

Parameters in MerQur

Date (optional) and Series (returns)

Concept: Because GARCH models the effect of yesterday's shock on today's variance, the observations must be in the correct time order. The series is the variable whose volatility is modelled, usually the return on a price.

Everyday example: If the exchange-rate data were recorded from the newest day to the oldest, selecting the date puts the days in the right order. The firm can fit the same model separately to the dollar and the euro rates; the two may show different persistence.

What it does: Sorts the observations by the date/order column (file order if empty) and selects the column to be modelled.

Effect on the result: Without a date, a reverse-ordered file is modelled backwards.

On the validation data: In the reversed sp500 file, selecting the date gives the same model as the ordered file. DAX and SMI log returns give different models (persistence 0.956 and 0.855).

Transformation (None (series already returns) / Log return × 100 / Percentage change)

Concept: Price levels are non-stationary and unsuitable for a volatility model, so they are converted into the relative change between successive days (the return). The log return × 100 and the percentage change are nearly the same for small changes; the first observation is lost because it has no predecessor.

Everyday example: If the rate rises from 30.00 to 30.60, the percentage change is 2.0% and the log return × 100 is about 1.98. The model is fitted to these daily changes, not to the price itself.

What it does: Converts a price level into returns: 100·Δln x or 100·(x_t/x_{t-1} - 1) (the first observation is dropped).

Effect on the result: Parameters and likelihood. Required when a price or index level is selected.

On the validation data: DAX log return: ω = 0.0476, α = 0.0685, β = 0.8876, LL = -2594.80; percentage change: ω = 0.0440, α = 0.0679, β = 0.8912, LL = -2587.92.

→ Log return

Variance model (GARCH (sGARCH) / EGARCH / GJR-GARCH)

Concept: Standard GARCH assumes positive and negative shocks raise volatility equally. EGARCH, by modelling the logarithm of the variance, and GJR-GARCH, by adding an extra term for negative shocks, capture the asymmetric effect whereby bad news raises volatility more than good news.

Everyday example: If the exchange rate stays turbulent for days after the local currency falls sharply, but calms down at once after it strengthens, the effect is asymmetric. The extra coefficient of GJR-GARCH then comes out significant.

What it does: sGARCH (Bollerslev 1986), EGARCH (Nelson 1991; ln σ² equation, sign α and size γ), GJR (Glosten et al. 1993; γ added for negative shocks).

Effect on the result: All parameters and the volatility dynamics.

On the validation data: dmbp LL: sGARCH -1106.587, EGARCH -1102.258 (β = 0.9125, γ = 0.3328), GJR -1106.084 (γ = 0.0283).

→ Conditional variance (volatility) and GARCH orders

ARCH order q (α) and GARCH order p (β)

Concept: q is how many days of past squared shocks the model uses (α terms), and p how many days of past conditional variance (β terms). For most financial series GARCH(1,1) is enough; higher orders should be added only if the information criteria improve clearly.

Everyday example: GARCH(1,1) says 'today's risk depends on yesterday's surprise and yesterday's risk'. If turbulence in the exchange rate is thought to depend separately on shocks two days back as well, q = 2 can be tried.

What it does: Number of α (lagged squared shock) and β (lagged conditional variance) terms. The model label is written "GARCH(p,q)" (p = β, q = α).

Effect on the result: Number of parameters and likelihood. At boundary solutions (α_i = 0) the standard errors are written "-" and a warning is given.

On the validation data: dmbp: q = 2 gives α2 = 0 (boundary), LL = -1106.947; p = 2 gives β1 = 0.490, β2 = 0.297, LL = -1104.329; p = 3 gives LL = -1099.064.

→ Conditional variance (volatility) and GARCH orders

Mean equation (Constant / ARMA(1,1) / Zero mean)

Concept: Alongside the variance model, a model for the mean of the return is also fitted: a constant mean, an ARMA(1,1) that captures slight dependence between successive returns, or a zero mean. The shocks (ε) are the departures of the return from this mean.

Everyday example: If the rate rises by 0.05% a day on average over the long run, a constant mean captures this drift so that the steady rise is not counted as a 'shock'. If the daily changes fluctuate around zero in practice, a zero mean is enough.

What it does: The mean model.

Effect on the result: Mean parameters and residuals.

On the validation data: dmbp ARMA(1,1): ar1 = -0.410, ma1 = 0.465, LL = -1103.890; zero mean: LL = -1106.854.

→ Deterministic terms (constant, trend)

Error distribution (Normal / Student-t)

Concept: The error distribution sets which distribution the standardised shocks come from. In financial returns extreme values occur far more often than the normal distribution predicts (heavy tails); the Student-t distribution accounts for this through its degrees of freedom (ν), with smaller ν meaning heavier tails.

Everyday example: If, under a normal distribution, a 3% daily jump should happen once a decade but the firm's data show it several times a year, the tails are heavy. A model with a Student-t distribution gives a more realistic probability for such days.

What it does: Normal or unit-variance Student-t (ν estimated). The forecast interval uses the quantile of the model distribution.

Effect on the result: Parameters, likelihood and intervals.

On the validation data: dmbp with t: LL = -989.830, ν = 4.356, persistence 0.999 (at the boundary; an "IGARCH boundary" warning is given). The 99% interval quantile is 3.204 (normal 2.576).

→ Error distribution (normal, Student-t)

Volatility forecast horizon

Concept: Sets how many days ahead the model forecasts volatility. Over a long horizon the forecast volatility approaches the series' long-run (unconditional) volatility at a speed set by the persistence.

Everyday example: After a turbulent day the volatility forecast for tomorrow is high, but the forecast for a month ahead is close to the exchange rate's 'normal' volatility. The firm looks at the horizon that matches its payment date.

What it does: σ_{T+h} and the mean forecast (by the conditional variance recursion).

Effect on the result: Number of forecasts.

On the validation data: dmbp σ T+1 = 0.3835; T+10 = 0.4286; T+250 = 0.5141 (converges to the unconditional σ = 0.5148).

→ Conditional variance (volatility) and GARCH orders

Confidence Level (%90 / %95 / %99 / Other, custom value)

Concept: Sets the confidence level of the interval given for the return forecast; the interval is built from the forecast mean and the forecast volatility. When volatility is high, the interval at the same confidence level widens too.

Everyday example: In a calm period the 95% interval for tomorrow's exchange-rate change might be ±0.3%, but in a turbulent one it can rise to ±3%. The firm reads from the interval's limits how far the rate could move against it tomorrow.

What it does: Only the forecast interval: mean ± z·σ.

Effect on the result: Does not change the coefficients or diagnostics.

On the validation data: In the normal model the intervals are given as mean ± z·σ using the standard normal quantile.

→ Confidence level and confidence interval

Ljung-Box lag and ARCH-LM lag

Concept: If the model is adequate, the standardised residuals (z) and their squares (z²) should be free of autocorrelation. Ljung-Box on z checks the adequacy of the mean equation, while Ljung-Box on z² and ARCH-LM check the variance equation; the lags are how many days back the tests look.

Everyday example: If z² is still autocorrelated after the model is fitted, part of the volatility clustering has not been captured and a higher-order or asymmetric model is tried. Ten lags look back about two trading weeks.

What it does: Ljung-Box Q(k) for the standardised residuals z and z² (fitdf = ARMA order for z); Engle ARCH-LM (n R²) on z.

Effect on the result: Diagnostic statistics.

On the validation data: dmbp LB z/z²: 5.06 / 2.50 for k = 1, 10.12 / 9.06 for 10, 19.29 / 17.50 for 20; ARCH-LM q = 1, 5, 12: 2.50, 4.21, 9.77.

→ Ljung-Box (portmanteau) test

Variance initialization (Sample variance (rugarch) / Backcast (FCP benchmark))

Concept: The GARCH recursion needs the variance of the first day; as this value lies outside the data, it is set by a starting rule. In long series the effect of this choice fades quickly; in short series it can change the coefficients slightly.

Everyday example: If the firm has only a few months of exchange-rate data, whether the first day's variance is set to the sample variance or by backcasting can change the results a little. With several years of data the two choices give almost the same result.

What it does: The first conditional variance: default h_1 = s² (sample variance); the "backcast" start h_1 = ω + (α + β)s² (the start of the FCP benchmark study).

Effect on the result: Parameters and likelihood.

On the validation data: dmbp backcast: μ = -0.0061904, ω = 0.0107614, α = 0.153134, β = 0.805974, LL = -1106.608.

→ Conditional variance (volatility) and GARCH orders

Example on real data

Data: EuStockMarkets DAX 1991-1998, R datasets

Variables: series DAX (daily closing index), transformation log return × 100

  • Question: Is there volatility clustering in DAX daily log returns (x 100, n = 1859), and how persistent is it?
  • Model: GARCH(1,1), constant mean, normal errors; μ = 0.065 (SE 0.022), ω = 0.048 (SE 0.013), α1 = 0.068 (SE 0.015), β1 = 0.888 (SE 0.024); all p < .01 (p = .163 for ω with the robust SE).
  • Persistence: α1 + β1 = 0.956; half-life 15.41 days, unconditional variance 1.082 (σ = 1.04).
  • Fit: log L = -2594.80, AIC/n = 2.796, BIC/n = 2.808.
  • Diagnostics: Ljung-Box Q(10) = 3.20 (p = .976) for the standardised residuals and Q(10) = 0.89 (p = 1.000) for their squares, ARCH-LM(5) = 0.61 (p = .988); the volatility structure is captured. Jarque-Bera = 13383.2 (p < .001; kurtosis 15.95): a Student-t distribution is recommended for the heavy tails.
  • APA: GARCH(1,1) model, α1 = 0.07, β1 = 0.89, persistence = 0.956, LL = -2594.80; Q²(10) = 0.89, p = 1.000.

Related concepts

Conditional variance (volatility) and GARCH ordersError distribution (normal, Student-t)Log returnForecast horizon and forecast intervalDeterministic terms (constant, trend)Ljung-Box (portmanteau) testPrediction intervalResidual plots and residual diagnostics

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.