AI GDP nowcasts trail human-adjusted forecasts by 0.01 points
Since 2023, the Czech National Bank's human-adjusted GDP nowcasts have recorded an error of 0.21 percentage points, against Random Forest's 0.22. The bank will update model weights quarterly using recent performance.
Human judgement retains its edge
The Czech National Bank (CNB) finds that its human-adjusted forecast narrowly beats Random Forest against flash GDP estimates; the weighted model average ranks third.
Accuracy is measured by root mean square error, not mean absolute error.
The toolkit includes the Rushin index, a dynamic factor model, LASSO, Ridge Regression, Elastic Net and Support Vector Machine alongside Random Forest.
The bank dropped an alternative dynamic factor model as its contribution diminished.
Previously fixed weights reflected accuracy after 2020; future weights will use performance since 2023 and update quarterly.
In renewed turbulence, the bank will also consider past crisis performance.
Experts adjust forecasts for events absent from, or slow to enter, the data.
Gains beyond calmer economic conditions
The CNB forecast beat an AR(1) benchmark by about 15 percent in 2015–2019, 80 percent in 2020–2022 and almost 35 percent since 2023.
Its error fell from below 0.7 percentage points before the pandemic to around 0.2 after exceeding 0.9 during the crisis period.
Rushin tracked the GDP fall of more than 8 percent in Q2 2020 and subsequent rebound above 7 percent.
Evaluation uses predictions roughly two weeks after quarter-end, comparing them with flash, first, second and latest GDP estimates.
Near parity, not replacement
Random Forest nearly matches the human-adjusted forecast, but the CNB reports no machine-learning lead yet.
The bank retains multiple models because their performance varies across economic conditions.
Quarterly reweighting changes how estimates are combined, while expert judgement remains part of the process.
Source: How accurate are AI models at nowcasting GDP?
IN: