Improving Business Cycle Forecasts' Accuracy - What Can We Learn from Past Errors?

Abstract

This paper addresses the question whether forecasters could have been able to produce better forecasts by using the available information more efficiently (informational efficiency of forecast). It is tested whether forecast errors covariate with indicators such as survey results, monetary data, business cycle indicators, or financial data. Because of the short sampling period and data problems, a non parametric ranked sign test is applied. The analysis is carried out for GDP and its main components. The study differentiates between two types of errors: Type I error occurs when forecasters neglect the information provided by an indicator. As type II error a situation is labelled in which forecasters have given too much weight to an indicator. In a number of cases forecast errors and the indicators are correlated, though mostly at a rather low level of significance. In most cases type I errors have been found. Additional tests reveal that there is little evidence of institution specific as well as forecast horizon specific effects. In many cases, co-variations found for GDP are not refected in one of the expenditure side components et vice versa.