Model Accuracy: Average, Over Time, and Its Relationship with Dispersion

Abstract

Many times economic forecasts are used as arguments for a change in a monetary policy or a basis for decisions by business people. The majority of modelers do not release a measure of accuracy of the predictions. This leaves the question of how good have the predictions been in the past. It was found that prediction intervals using a 95 percent level of confidence were large enough to put their use into question for three macroeconomic variables: the growth in real GNP, inflation, and the unemployment rate. The prediction intervals calculated were over a fourteen year period. This left the question of when the errors were the greatest. It was hypothesized that the errors in prediction would be large for forecasts made in peaks and troughs. This is due to the fact that during these times economic activity does not continue to increase or decrease, but changes direction faster than the trend. It was discovered that the errors of predictions made in peaks or troughs were not necessarily greater than the mean error and that large errors could have been caused by a range of events. Finally, the relationship between the dispersion of forecasts and the associated error was considered. If a relationship were found, the people who use the forecasts could know the level of confidence which should be placed in a prediction if a given dispersion were witnessed. Unfortunately, it was found that there is no clear relationship between the two.