The Economic Logic of Large Language Models

Abstract

Unlike narrow statistical models, large language models (LLMs) can extrapolate patterns across disparate domains of knowledge. We apply a logical interpretation framework called Model Fingerprint to quantify the precise macroeconomic reasoning of LLMs. We find that they identify intuitive economic relationships, including conditionalities that diverge strongly from machine learning approaches applied narrowly to economic data. LLMs were able to infer statistically significant differences in positive versus negative contemporaneous economic growth from five other economic inputs, outperforming logistic regression and random forest models trained on corresponding historical data.

Keywords: Large Language Models, LLM, GPT, Machine Learning, Ensemble Models, Model Interpretability, Economic Growth