Re: [Gretl-devel] OLS Forecasting

On Mon, 9 Feb 2009, Sven Schreiber wrote: [Warning: long discursive reply!] Yes, I will admit that there are diminishing returns to pursuing our current design strategy, and that we may have something to learn from PcGive. The thing is that for historical reasons -- but not _just_ historical reasons, because it works very nicely, up to a point -- we have a very simple basic mechanism for specifying a regression in gretl. Namely, the "list": an array of integers. The first value tells you how many values follow, and the remaining values are, first, the dependent variable (i.e. the position of that variable within a two-dimensional array of series), then the independent variables (again, as indices into the data array). It's easy to extend this pattern to estimators such as two-stage least squares (which requires a distinct list of instruments) by defining a "list separator" integer value; and we have a well honed suite of functions, in gretl_list.c, that will pick apart or splice sub-lists defined in this way. And it's not difficult to extend the mechanism further, by passing an additional integer (not part of the list itself) to certain estimators, to specify a global lag order (as in VARs and VECMs). It's maybe worth remarking that, in the case of GUI dialogs that present a more "structured" appearance with regard to lags, what's going on in the background (for the most part) is that these structured representations are converted into "list" form for internal consumption. That is, the required lags are generated as additional series in the data array, and a list is composed containing the indices of the lagged variables (as if they were just "any old" variables). Now, if we had designed from scratch a program for analysis of time series it's unlikely we would have chosen such a representation. The "list", although it can contain several sub-lists, is inherently one-dimensional, and so is not a very convenient mechanism for specifying tuples of the form (variable, lag specification). Note that the design issue here is a matter of relative convenience and clarity, not possibility/impossibility. The Turing machine has a linear tape, yet can compute any computable function. We could certainly pack fully explicit lag-polynomial specifications into a "list" (perhaps with the use of more than one "separator" integer). And this may be worth considering. But maybe the time is approaching when we should think about substituting a more structured data type for the regression "list". This is a bit daunting, since the gretl code base contains hundreds if not thousands of calls to regression-related functions that currently expect a list argument, but it's doable. Allin.