Generalization of information, Granger causality and forecasting

Purpose This paper aims to analyze forecasting problems from the perspective of information extraction. Circumstances are studied under which the forecast of an economic variable from one domain (country, industry, market segment) should rely on information regarding the same type of variable from another domain even if the two variables are not causally linked. It is shown that Granger causality linking variables from different domains is the rule and should be exploited for forecasting. Design/methodology/approach This paper applies information economics, in particular the study of rational information extraction, to shed light on the debate on causality and forecasting. Findings It is shown that the rational generalization of information across domains can lead to effects that are hard to square with economic intuition but worth considering for forecasting. Information from one domain is shown to affect that from another domain if there is at least one common factor affecting both domains, which is not (or not yet) observed when a forecast has to be made. The analysis suggests the theoretical possibility that the direction of such effects across domains can be counter-intuitive. In time-series econometrics, such effects will show up in estimated coefficients with the “wrong” sign. Practical implications This study helps forecasters by indicating a wider set of variables relevant for prediction. The analysis offers a theoretical basis for using lagged values from the type of variable to be forecast but from another domain. For example, when forecasting the bond risk spread in one country, introducing in the time-series model the lagged value of the risk spread from another country is suggested. Two empirical examples illustrate this principle for specifying models for prediction. The application to risk spreads and inflation rates illustrates the principles of the approach suggested here which is widely applicable. Originality/value The present study builds on a probability theoretic analysis to inform the specification of time-series forecasting models.