The HAC Emperor Has No Clothes: Part 3

If you've been reading this blog for a while, you know I'm no fan of nonparametric HAC time-series regression.  See here and here.  Instead just do parametric ARMA disturbances or dynamic regressions, select orders using AIC etc., estimate by feasible GLS or MLE.  It's trivial to implement (even in 1985, say) and basically as good as it gets for efficient estimation and inference (and of course it too is nonparametric from a sieve perspective).  Plus you can exploit the parametrically-captured serial correlation for improved time-series prediction.

A new paper by Dick Baillie, George Kapetanios and Kun Ho Kim, "Practical Approaches to Achieve Robust Inference in Time Series Regressions," brings all this into even sharper focus.  Their beautiful Figure 1, reproduced below, says it all.  The black dot is feasible GLS; the reds and blues are the standard nonparametric competitors.  Presumably the paper will be posted soon.  A related much earlier paper is here.