Sweden's Riksbank has published an Economic Commentary showing how combining Statistics Sweden's preliminary gross domestic product (GDP) indicator with an existing forecast can improve growth estimates before the National Accounts (NA) are released. The indicator becomes more informative as additional months of a quarter become available, but contains noise and should not be treated as an actual NA outcome. The proposed model adjusts both sources for their historical tendency to overestimate or underestimate growth, then combines them using weights based on their historical accuracy and the correlation between their errors. The indicator receives a weight of 2% when one month of data is available, rising to 22% with two months and 73% with the full quarter. Even with all three months available, the model retains some weight on the existing forecast rather than relying entirely on the indicator. Across 24 quarters from the third quarter of 2020 to the second quarter of 2026, the model's root mean square error was 39% lower than that of the Riksbank's forecast when indicator data covered the full quarter. Performance was measured against the first NA release, not final GDP figures, and the improvement was statistically significant only with all three months available. Because the model's weights were estimated using the entire sample, a separate evaluation used only information available at each forecast date. That test showed a smaller reduction of 22% with a full quarter of indicator data. The short sample means the estimated weights and forecasting gains may change as more observations become available.
Sweden's Riksbank study finds combining GDP indicator and forecasts lowers forecast error by 39% in sample
Sweden's Riksbank has published analysis showing that weighting Statistics Sweden's preliminary GDP indicator together with an existing forecast can improve growth estimates, rather than treating the indicator as an actual outcome. With all three months of a quarter available, the model reduced root mean square forecast error by 39% in the full sample and 22% in a separate test using only information available at each forecast date. The results draw on a short data history, so the estimated gains may change as more observations become available.