one-step forecasts is, of course, simply the analogy of many in-sample tests of model specification. In this context, it is amusing that much recent applied econometric work has, in effect, exploited fitted models to make inference about economic systems, as the forecast horizon approaches injinity! This work includes the literature on unit roots, cointegration, persistence of shocks, and unobserved components.
The second 'surprise' in the ClementslHendry thesis is the recommendation that, however many series are predicted, evaluation should be based on just a single function-the generalized second moment-of the one-step forecast errors. Again, as is clear from their paper, this is merely a post-sample check on model adequacy. Suppose Hendry and I were invited to forecast two time series-monthly sales of coconuts in Australia and annual rainfall in the Andes. Hendry's greater econometric expertise would doubtless give him an advantage in the first case, and I am prepared to concede defeat. However, I might-perhaps by luck-arrive at a superior predictor of Andean rainfall. I would want this to be judged separately, rather than merged with the Australian coconuts. Is this an unreasonable view, even to an Andean umbrella salesperson?
In their conclusions, Clements and Hendry invite us to consider the mind-set of a British Chancellor of the Exchequer. I find this difficult, only partly because I have never met one. Still, I am astonished, and not particularly reassured, by the notion that he, or some future she, would care about the model from which forecasts are generated. I can see that the Chancellor would care that whatever is to be predicted, and however far ahead it is to be predicted, the forecasts be as accurate as possible. I wonder how sanguine are some recent Chancellors on the question of whether econometricians are achieving near-truth in their models?