A0221
Title: LLMs vs econometric models for nowcasting GDP growth: A practitioner's view
Authors: Julien Andre - University Paris Dauphine/Banque de France (France)
Marie Bessec - University Paris Dauphine (France) [presenting]
Zachary Goulby - University Paris Dauphine PSL (France)
Abstract: The performance of large language models (LLMs) is evaluated in nowcasting French GDP growth, comparing them with the econometric models currently used by the Banque de France. Using only prompt-based queries without external data or fine-tuning, it is assessed whether general-purpose LLMs such as ChatGPT, Gemini, and Claude can serve as effective forecasting tools. While econometric models consistently outperform LLMs during normal periods, the latter show a notable advantage in capturing exceptional events such as the COVID-19 pandemic. The sensitivity of LLM forecasts is also examined to prompt language, design, and model version, and a confidence index and recession probability derived from LLM responses are introduced. There is no strong evidence of information leakage, and robustness checks confirm the findings across various model versions and temperature settings. A fair in-sample comparison reinforces the relative strength of econometric models in normal conditions. Overall, the results suggest that while standard LLMs are not yet ready to replace traditional models in routine forecasting, they can provide complementary insights, particularly in periods of structural change or heightened uncertainty. These results apply to the most popular LLMs without external data and fine-tuning.
