Combining GDP signals cuts forecast error by 39 percent in sample
Viking Waldén's Riksbank analysis combines Sweden's GDP indicator with existing forecasts, reducing root mean square error by 39 percent over Q3 2020–Q2 2026 when full-quarter data are available. The reduction falls to 22 percent in an out-of-sample test.
Three months sharpen the signal
Waldén's model combines the Riksbank's GDP forecast with Statistics Sweden's monthly GDP indicator to predict the first National Accounts release.
It corrects both inputs for historical bias and weights them using forecast-error variances and covariance.
Across 24 quarters from Q3 2020 to Q2 2026, the indicator receives weights of 0.02, 0.22 and 0.73 as one, two and three months become available.
The model's root mean square error falls from 0.79 to 0.75 and 0.50 percentage points, respectively, against 0.82 for the Riksbank forecast.
The full-quarter indicator alone records 0.62. Only the full-quarter improvement over the Riksbank forecast is statistically significant, with a p-value of 0.01.
Real-time testing trims the gain
Full-sample estimates use later observations to calibrate earlier forecasts.
An expanding-window test instead starts with ten observations and evaluates 14 quarters from Q1 2023 to Q2 2026.
Its root mean square errors are 0.84, 0.78 and 0.63 percentage points with one, two and three months of indicator data.
Relative to the Riksbank forecast, these are 1.04, 0.96 and 0.78. Excluding 2020–2021 changes the main results little.
The indicator remains preliminary: some investment and government consumption data are unavailable when it is compiled.
An estimate, not an outcome
The full-quarter GDP indicator improves forecasting but does not replace the National Accounts.
The smaller out-of-sample gain shows that full-sample performance overstates predictive power.
With only 24 quarters available, the estimated weights and improvements remain sensitive to additional observations.