Illuminating insights into economic models of distinction

The content of these three essays comprises the material from Clive Granger's Alfred Marshall lectures of 1998. Wide ranging in scope, they condense a large amount of material into a short book.

His theme is the process of modelling, from specification to evaluation, and he states a forthright and openly pragmatic view. He believes the ultimate yardstick for judging a model should be the economic value of using the output of the model in real decisions. Different models are appropriate for different purposes, and we cannot judge them without knowing their purpose or intended use.

Granger is critical of widespread practice, particularly in cross-sectional empirical work. He expresses himself to be "amazed that some methodologists appear to be content with economics providing an 'explanation' for what has occurred in the past". Explanations are generally not unique, and he believes it is impossible to distinguish which is correct. Many applied workers would see their role as seeking ways to test the implications of economic theory against data drawn from appropriate contexts. Granger's tone suggests a distrust of abstract economic theories - they are "naive" and typically too vague to allow them to be matched to data, particularly in dynamic aspects. In a telling example, he takes results in one familiar piece of work and shows how lack of clarity in specifying units of time relevant to the theory undermines the interpretation of results.

For evaluating models, he argues that in-sample fit is rendered ineffectual as a guide to model adequacy by the process of generating models. His preference is for attention to be paid to predictions on data not used in model design.

The final chapter, which concentrates on evaluation of forecasts, is the most rewarding. If we think seriously about the purposes for which models are used, we shall admit that losses based on forecast inaccuracy do not take the simple forms that underlie usual criteria for evaluation. Granger points to results establishing how forecast assessment generalises to cases of other specific loss patterns. In general, however, if model comparisons are to be available for all potential users, the complete predictive distribution of the forecasts is needed and simple summary statistics will not do. An elegant conceptual framework begins to be worked out for evaluations that anticipate how decisions might be reached on the basis of comprehensive reporting of forecast properties.

These essays state strong opinions based on what may seem an unduly restrictive view of how economic modelling can usefully contribute to knowledge, but they contain a lot of wisdom and illuminating insights.

Ian Preston is senior lecturer in economics, University College London.