Assess the EGARCH Forecast Bias Using Simulations

This example shows how to simulate an EGARCH process. Simulation-based forecasts are compared to minimum mean square error (MMSE) forecasts, showing the bias in MMSE forecasting of EGARCH processes.

Step 1. Specify an EGARCH model.

Specify an EGARCH(1,1) process with constant $\kappa = 0.01$, GARCH coefficient $\gamma_1 = 0.7$, ARCH coefficient $\alpha_1 = 0.3$ and leverage coefficient $\xi_1 = -0.1$.

Mdl = egarch('Constant',0.01,'GARCH',0.7,...
    'ARCH',0.3,'Leverage',-0.1)
Mdl = 

    EGARCH(1,1) Conditional Variance Model:
    -----------------------------------------  
    Distribution: Name = 'Gaussian'
               P: 1
               Q: 1
        Constant: 0.01
           GARCH: {0.7} at Lags [1]
            ARCH: {0.3} at Lags [1]
        Leverage: {-0.1} at Lags [1]

Step 2. Simulate one realization.

Simulate one realization of length 50 from the EGARCH conditional variance process and corresponding innovation series.

rng default; % For reproducibility

[v,y] = simulate(Mdl,50);

figure
subplot(2,1,1)
plot(v)
xlim([0,50])
title('Conditional Variance Process')

subplot(2,1,2)
plot(y)
xlim([0,50])
title('Innovations Series')

Step 3. Simulate multiple realizations.

Using the conditional variance and innovation realizations generated in Step 2 as presample data, simulate 5000 realizations of the EGARCH process for 50 future time steps. Plot the simulation mean of the forecasted conditional variance process.

rng default; % For reproducibility
[Vsim,Ysim] = simulate(Mdl,50,'NumPaths',5000,...
                       'E0',y,'V0',v);

figure
plot(v,'k')
hold on
plot(51:100,Vsim,'Color',[.85,.85,.85])
xlim([0,100])
h = plot(51:100,mean(Vsim,2),'k--','LineWidth',2);
title('Simulated Conditional Variance Process')
legend(h,'Simulation Mean','Location','NorthWest')
hold off

Step 4: Compare simulated and MMSE conditional variance forecasts.

Compare the simulation mean variance, the MMSE variance forecast, and the exponentiated, theoretical unconditional log variance.

The exponentiated, theoretical unconditional log variance for the specified EGARCH(1,1) model is

$$\sigma _\varepsilon ^2 = \exp \left\{ {\frac{\kappa }{{\left( {1 - {\gamma _1}} \right)}}} \right\} = \exp \left\{ {\frac{{0.01}}{{(1 - 0.7)}}} \right\} = 1.0339.$$

sim = mean(Vsim,2);
fcast = forecast(Mdl,50,'Y0',y,'V0',v);
sig2 = exp(0.01/(1-0.7));

figure
plot(sim,':','LineWidth',2)
hold on
plot(fcast,'r','LineWidth',2)
plot(ones(50,1)*sig2,'k--','LineWidth',1.5)
legend('Simulated','MMSE','Theoretical')
title('Unconditional Variance Comparisons')
hold off

The MMSE and exponentiated, theoretical log variance are biased relative to the unconditional variance (by about 4%) because by Jensen's inequality,

$$E(\sigma _t^2) \ge \exp \left\{ {E(\log \sigma _t^2)} \right\}.$$

Step 5. Compare simulated and MMSE log conditional variance forecasts.

Compare the simulation mean log variance, the log MMSE variance forecast, and the theoretical, unconditional log variance.

logsim = mean(log(Vsim),2);
logsig2 = 0.01/(1-0.7);

figure
plot(logsim,':','LineWidth',2)
hold on
plot(log(fcast),'r','LineWidth',2)
plot(ones(50,1)*logsig2,'k--','LineWidth',1.5)
legend('Simulated','MMSE','Theoretical')
title('Unconditional Log Variance Comparisons')
hold off

The MMSE forecast of the unconditional log variance is unbiased.

See Also

| |

Related Examples

More About

Was this topic helpful?