An Instrumental Variables Approach to Testing Forecast Efficiency

ABSTRACT We study a specific form of forecast efficiency that requires forecast errors to be unpredictable from forecast revisions. One approach aggregates forecasts and estimates an aggregated efficiency regression, while another estimates the relationship by running separate regressions for each individual and then aggregating. We demonstrate that both estimators can be asymptotically biased in the presence of public noise. To address these biases, we propose instrumenting forecast revisions with past forecast errors. The Anderson–Rubin likelihood ratio test can be applied to test for forecast efficiency and remains robust even in the presence of weak instrumental variables. Applications of the test to the US Survey of Professional Forecasters clearly reveal experts' underreaction to news in their macroeconomic expectations.