A Bayesian methodology for averaging predictions: Application to BCRA's market expectations survey

Abstract

The Central Bank of Argentina (BCRA) publishes the Market Expectations Survey (REM) monthly, which summarizes the economic projections and predictions made by a group of economic analysts and consultants. The BCRA discloses only the main aggregate statistics of the sample, such as the median, average and standard deviation. The logic for using these statistics is that all participants should be weighted similarly. If some consultants are thought to have better underlying models than others, the effectiveness of aggregate forecasts can be substantially improved by prioritizing the predictions of those who have historically forecasted better. Even without knowing in detail the models used, there is information on the predictions made by them in the past. A method that weights such historical performance should lead to a better aggregate average. In this work, a Bayesian method is developed that allows calculating such weights. The aggregate average resulting from Bayesian weights provides statistically better predictions than the arithmetic mean, median, and other commonly used methods. In particular, the developed method more effectively detects trend changes in the projections. The published aggregate predictions of the REM provide useful information, not only for monetary and economic policy decisions, but also for consumption and investment decisions. Therefore, improving these predictions benefits all agents in the economy.

JEL classification: C11 ; E52