We turned timestamped Federal Reserve Beige Book PDFs into forecasting data, then trained a compact model to predict changes in regional economic conditions. The model beats GPT-5 by 22% on prediction error and 84% on calibration using only the PDFs as input.
What we did
- Read timestamped Fed Beige Book PDFs and generated district-level economic forecasting questions from each edition.
- Resolved each question from the next Beige Book, using future text as the label rather than manual annotation.
- Packaged the full PDF-to-forecaster workflow as a runnable SDK notebook.
Read more
Primary artifacts for this case study.