Sweden's Riksbank staff memo finds AI forecasting models outperform traditional approaches for Swedish GDP and inflation

Sweden's Riksbank published a Staff Memo, “AI-based forecasting in Sweden”, assessing whether machine learning models improve forecasts for Sweden’s monthly GDP indicator and CPI inflation. Comparing random forests and neural networks with standard time-series benchmarks and factor models, the authors find that AI-based models deliver more accurate forecasts, with the strongest gains for CPI inflation. The study uses a dataset of 120 Swedish and foreign macroeconomic and financial variables and evaluates out-of-sample performance over January 2017 to February 2024 across 1-, 2-, 6- and 12-month horizons. Random forests provide the most accurate inflation forecasts at all horizons, improving forecast accuracy by around 10–40 percent relative to a random-walk benchmark, while gains for the GDP indicator are smaller and more similar across models (autoregressive, factor and random forest models are around 15–30 percent more accurate than the random walk). Diebold-Mariano tests indicate the random-forest advantage for CPI inflation is often statistically significant versus factor models, whereas differences for GDP forecasts are not statistically significant; the memo also concludes that model nonlinearity is more important for forecast performance than simply expanding the information set.