Comparative Efficiency of Maximum Likelihood and ex ante Reduced Form for Forecasting: Study of a Simple Model

R. C. GEARY 

The famous issue of maximum likelihood (ML) versus least squares (LS) in the solution of a behavioristic equation system flares up from time to time, but it seems to be as yet unresolved. 
Accordingly, since the writer is about to embark on a possibly large model for his own country - based on time series - he himself has to face the issue now. The present investigation, though based on a very special model, leaves him under the impression that the case for ex ante reduced form (RF) is better than he had supposed. Of course, "it all depends on what one wants the model for" - to quote the familiar cliché. One objective that the wriiter does not have in mind is individual coefficient estimation. Some years ago he vehemently asserted that in multivariate regression (and, a fortiori, in equation systems) individual coefficients are meaningless: the only coefficients possibly economically significant are those of simple regression. The writer is not aware af any 
serious attempt to rebut his views; nonetheless, economic interpretations of individual coefficients (usually interpreted as "'elasticities" or the like), with their implicitly untenable ceteris parabus 
assumption, are still widespread.