2
$\begingroup$

Diebold & Mariano suggest using Newey-West standard errors to correct for autocorrelation and heteroscedasticity in the error terms when comparing forecast accuracy.

However, if I understand correctly, they also assume dependency only for (k-1) day ahead forecast errors.

As such, am I correct in assuming that I do not need to use Newey-West standard errors when looking at 1-day ahead forecasts?

$\endgroup$
2
  • $\begingroup$ The implementation code of the test in R package "forecast" function dm.test is quite easy to read; there you see that the variance estimation accounts for autocorrelation up to $h-1$. $\endgroup$ Aug 12 '16 at 9:45
  • $\begingroup$ I was going through my old answers and noticed this one was not accepted. Do you perhaps need further clarification? $\endgroup$ Mar 17 '17 at 11:47
3
$\begingroup$

Autocorrelation
$h$-step-ahead forecast errors for $h>1$ are autocorrelated by construction, because the forecasts cover overlapping time periods. This holds even for "optimal" forecasts. Meanwhile, 1-step-ahead forecast errors need not be autocorrelated, at least for "optimal" forecasts.

However, they still might be autocorrelated for "suboptimal" forecasts. E.g. if the forecast for $t+1$ is just the actual value as of $t-k$ for some $k>0$, you effectively get the same overlap and thus autocorrelation in forecast errors as with $h$-step-ahead forecasts. Of course, this need not be a sensible forecast, but it serves as an example.

Using autocorrelation-robust standard errors for 1-step-ahead forecasts in the Diebold-Mariano test is thus a conservative, "safe" choice.

Heteroskedasticity
Regarding heteroskedasticity, neither Diebold & Mariano (1995) nor Diebold (2015) mentions it, so probably there is nothing special about it in the context of the Diebold-Mariano test. If there is reason to believe that the forecast error distribution had a varying variance in your sample, you could use the heteroskedasticity-robust standard errors.

References:

$\endgroup$
1
  • $\begingroup$ Now I started wondering myself what one should do with "suboptimal" forecasts. Does the fact that the 1-step-ahead forecast errors are autocorrelated effectively reduce the sample (as with optimal $h$-step-ahead forecasts) so that we have to use autocorrelation-robust variance estimator? Or do the errors carry independent information nevertheless and regular variance estimator should be used? $\endgroup$ Aug 12 '16 at 9:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.