Addressing Unemployment Rate Forecast Errors in Relation to the Business Cycle

In this paper we show that the prediction errors for unemployment vary over the business cycle. We initially use a macroeconomic data for the United States, because these data are available for a longer period than for the Netherlands. The dataset for the United States covers the 60s of the last century up to the COVID-crisis. We find that forecast errors are greatest in recession periods, but they are also relatively large in recovery periods. The forecasting errors are smaller in the periods between. All forecasting models show this error pattern, but interestingly, they don't all show it to the same extent. Some models perform relatively well during recessions, others during recovery periods, and still others during the periods in between. As a result, the choice of the best model depends on the weight that is assigned to prediction errors in recovery and recession periods. These findings are relevant to the forecasting models used by the CPB that support the bureau-wide unemployment forecast. We carried out the same analysis on Dutch data, which provided comparable results. For the Dutch unemployment estimate, too, the choice of the best model depends on the weights for recovery and recession periods.