1

Does anyone know how to implement Hodrick Standard errors in R? I could not find any package for it in R. Is anyone aware of the same or any open source code that implements it? I want to use Hodrick Standard errors after the estimation of e-GARCH and DCC-GARCH models in R. I am not sure how to approach this?

The code to estimate Hodrick standard errors after a regression model I use is:

kernHAC(r, kernel = "Truncated", bw = 1,
  prewhite = FALSE, adjust = FALSE, sandwich = FALSE)

Where r is a regression object.

However, the sandwich package does not support GARCH estimator. How do I estimate Hodrick standard errors after a GARCH regression?

Raghav Goyal
  • 138
  • 7
  • You mean the ones proposed in [Hodrick (1992)](https://www0.gsb.columbia.edu/faculty/rhodrick/dividendyields.pdf)? – Richard Hardy Nov 15 '20 at 19:33
  • @RichardHardy Yes – Raghav Goyal Nov 16 '20 at 03:21
  • While this is clearly not an answer to your question, I wonder why you would want to use such standard errors. Garch models are designed for uncorrelated, but not independent series - why would one then need HAC standard errors? – Christoph Hanck Nov 16 '20 at 13:33
  • @ChristophHanck, the problem might be overlapping observations. I am not entirely sure, but if that is the case, increasing the standard errors might be the easiest way out. See e.g. ["Covariance estimation of overlapping time series"](https://stats.stackexchange.com/questions/154979). I thought I had found a better way, but later stroke it through as I suspect it is not a valid approach. – Richard Hardy Nov 16 '20 at 13:55
  • @RichardHardy, yes, something along these lines seems possible. Although I would then probably try and go for some ARMA-GARCH type of approach to mop up serial correlation, if I understand the problem correctly. – Christoph Hanck Nov 16 '20 at 13:58
  • @ChristophHanck, just to make sure I understand you. Could using an ARMA error structure instead of a robust covariance estimator make sense for overlapping data? I know Diebold advocates ARMA errors over robust covariance estimators in general, but he has not proposed it for the Diebold-Mariano test in case of overlapping predictions. This makes me guess ARMA errors are invalid for overlapping data. I was never able to think this through clearly. I could post a separate question if there is a chance you would answer it. – Richard Hardy Nov 16 '20 at 14:20
  • @ChristophHanck Actually I was thinking on similar lines that GARCH standard errors should suffice since they take into account the heteroskedasticity and autocorrelation. However, the referees are asking me to report these standard errors. Hence, I am trying to reason out and see what can be done with HH standard errors. Also, I am dealing with overlapping time series in a DCC-GARCH setting. Would you suggest using HH standard errors in this setting? And, how should I approach them in R? – Raghav Goyal Nov 16 '20 at 14:48
  • Your concerns/thoughts are both relevant - unfortunately I am at a similar stage in terms of thinking it through... – Christoph Hanck Nov 16 '20 at 15:33

0 Answers0