Time-series foundation models versus professional forecasters
Foundation models are rising fast in the AI journals, and time-series versions now forecast series they were never trained on. Economists forecast inflation and growth with surveys and small models. Do the new models beat the Survey of Professional Forecasters in real time?
What prompted this
In the AI journals, "foundation models" had 27 papers in July–September 2026 against about 16 expected (see the AI dashboard). Since October 2025, 1,714 papers carry a foundation-model or large-language-model keyword, but only 5 of them also carry an inflation, macroeconomic or GDP forecasting or nowcasting keyword. "Inflation forecasting" appears on 31 papers (22 in economics), and "time series forecasting" on 148 (117 in AI journals).
Questions
- With real-time data only, do zero-shot time-series foundation models beat the median forecasts of the Survey of Professional Forecasters for inflation and GDP growth?
- Can pretraining leak the future into tests that look out of sample? Data vintages help check this.
- Where do the models fail: at turning points, at regime changes, or around data revisions?
Data that could help
Real-time data vintages (the Philadelphia Fed's real-time data set or ALFRED), the Survey of Professional Forecasters, and FRED-MD.
Next step
Run one open time-series foundation model on real-time CPI vintages for 2015–2025 and score it against the survey median.