Am 17.08.2017 um 21:27 schrieb Allin Cottrell: > On Wed, 16 Aug 2017, Ioannis A. Venetis wrote:
> I agree that "rolling" tends to imply a moving estimation window of > fixed length, and if "recursive" is the term used by forecasters for > what gretl does under the --rolling option, we should switch. This is > now in git (though "--rolling" is still accepted as an alias until > further notice). > > However, I'd like to register a mild objection. To my mind a recursive > process is one that loops back on itself (e.g. a function that calls > itself). In that light a "recursive" forecast would be one which > progressively replaces known values of regressors with forecasted values > as the horizon is lengthened. If I had to choose a single English word > to describe what gretl does under the rolling or recursive option I > guess I'd choose "sequential". I agree, and of course it's all a matter of convention. What you describe as the "ideal" meaning of recursive is also used in Diebold's forecasting texts. Actually you basically paraphrased him, here's a quote from his book: "Recursive estimation means beginning with a small sample of data, estimating a model, adding an observation and reestimating the model, and continuing in that fashion until the sample is exhausted.[FN9]" ...and FN9 says: "Strictly speaking, sequential might be a more descriptive adjective than recursive. Recursive updating refers to the fact that an estimate based on t + 1 observations can sometimes be computed simply by appropriately combining the old estimate based on t observations with the new observation. (This is possible, for example, with linear least-squares regression.) Recursive updating achieves a drastic reduction in computational requirements relative to complete reestimation of the model each time the sample is updated, which we might call "brute force updating." For our purposes, it's inconsequential whether we do recursive updating or brute force updating (and the speed of modern computers often makes brute force attractive); we use recursive estimation as a blanket term for any sequential estimation procedure, whether the computations are done by recursive or brute force techniques." The important thing would be to replace "rolling", but perhaps it's also a good time to be more accurate and indeed use "sequential". cheers, sven
