Econometric models are a widely used and powerful tool in macroeconomic analysis and forecasting. Admittedly, their acceptance by the academic community had some hard times during the seventies and eighties. The general decline in the reputation of macroeconomics, the Lucas and the Sims critique, as well as failures of the modelling community to make their models and their practice more transparent have left their marks. Closer looks at the ""critiques"", have, however, revealed their limited relevance, and the ""new/old macroeconomic consensus"" of the early 1990s seem to have restored much of the lost credibility.¤ Unfortunately this does not hold for the charge of missing transparency. More specific, it is criticised that (1) the numerous, often complex model relationships and the role and importance of endogenous/exogenous as well as other outside information are difficult to understand and (2) the models and especially their results are not free from bias, i. e. the models tell what their users want them to tell. This ""black box"" accusation is expressed in particular by members of the academic community, who, for various reasons, favor small, reductionist text book models. In contrast to this, policy makers and business economists value the informational content of large models. They are much less sceptical of these or overcome their doubts about the models by using others' models in addition to their own. The complaint about missing transparency is not new. Tinbergen's models in the 1930s were confronted with it. This complaint figured prominently in the macroeconometric model critique of the early seventies and it is still often heard today. The problem of ""reading"" or understanding macroeconometric models received some attention by employing verbal model descriptions, block diagrams, logical analysis (graphs, incidence matrices), model condensations, aggregated supply/demand curves, and implicit Phillips-curves. This, however, is not the case with respect to the generation of model forecasts. Of course, the aforementioned techniques facilitate also the understanding of the forecasting process, but the literature on this is still rare, and the problem is ignored in most econometric textbooks. The main explanation for all this is that there is not much effective user demand for model transparency. Consequently, the model industry sees no reason to improve the situation.¤ This paper deals with the possibilities of making forecast practice transparent by describing in detail the generation of a macroeconometric forecast. It should serve two main purposes. First, it structures the process of macroeconometric model forecasting using state of the art techniques. This should be a guide for producers as well as for users. Second, it provides users of these forecasts an understanding that model forecasts are not the result of ""black boxes"" but can be tracked down to hypotheses and assumptions. The model used is the model of the Rheinisch-Westfälisches Institut für Wirtschaftsforschung (¤ RWI), a medium sized, short term macroeconometric model for the (West) German economy. It has been regularly applied for forecasting since 1978. The forecast that is examined is the one that was made in autumn of 1996 for 1996 and 1997. The paper describes the various steps of forecast generation, shows their consequences for the final forecast and analyzes in particular the accuracy of this and some competing forecasts.¤ The next section displays some analytical foundations of econometric forecasting and error analysis. Section 3 briefly presents the model used and the macroeconomic situation in late 1996. The generation of the forecast is described in section 4, including its ex post assessment. The paper ends with a summary and conclusions. The subtitle of the paper announces a report from the trenches, hence the paper is rather dense and brief, and, equally important, it abstains from generalisations. Although the general framework is likely to hold for many macroeconometric model forecasts, the weight given to the various stages of the forecast production will vary with the purpose and the design of the model,
