A small economic forecaster trained from raw Fed PDFs beat GPT-5 (blog.lightningrod.ai)

A small economic forecasting model trained exclusively on the Federal Reserve's Beige Book PDFs has outperformed GPT-5, showing 22% better prediction accuracy and 84% improved calibration. The model, which received no manual annotations, utilized a process called Future as Label to transform narrative data from the Beige Books into binary forecasting questions related to economic conditions. By employing the innovative Foresight Learning technique, the model fine-tuned its predictions based on actual outcomes, allowing it to learn effective cause-and-effect reasoning in economic forecasting. This advancement holds significant implications for the AI/ML community, particularly in the realm of document-based prediction models. The success of this model highlights the limitations of existing large language models like GPT-5, which performed poorly despite having access to identical data. By demonstrating that automated data generation and outcome-based training can enhance predictive accuracy, this approach opens pathways for organizations in various sectors to develop tailored forecasting tools without the need for extensive labeled datasets. The authors have shared their methods as runnable code, encouraging other teams to apply these techniques to their own data.