Large language models for unstructured economic data processing: Enhancing GDP forecasting accuracy with heterogeneous textual information

Gross domestic product (GDP) forecasting is essential for timely economic policy, but traditional models struggle to utilize the forward-looking signals embedded in unstructured text. This paper bridges this gap by proposing a novel hybrid framework that leverages Large Language Models (LLMs) to process heterogeneous textual sources–including news, PMI commentary, central bank communication, and corporate disclosures into unified, interpretable economic indices. These indices are integrated with conventional macroeconomic indicators within both linear and non-linear forecasting models. Evaluated in a pseudo real-time setting across four distinct datasets, our approach demonstrates that LLM-augmented models consistently and significantly outperform structured-data benchmarks. The results show relative RMSE reductions of up to 29%, with the largest gains occurring during economic recessions. This underscores the particular value of textual narratives for capturing turning points. Our study provides policymakers and forecasters with a scalable, disciplined, and transparent methodology to exploit unstructured information for enhanced GDP nowcasting accuracy.