The Bank of England published Staff Working Paper No. 1,150, “Revealing economic facts: LLMs know more than they say”, examining whether the hidden states (embeddings) of open-source large language models (LLMs) can be used to estimate and impute economic and financial variables. The paper finds that a simple linear model trained on embeddings can outperform the same model’s natural-language text outputs when estimating county-level and firm-level statistics, and it demonstrates applications in data imputation and geographic “super-resolution”. The analysis covers regional datasets for the United States, United Kingdom, European Union and Germany (including variables such as unemployment, GDP per capita, income and life expectancy) and a dataset of United States listed firms (including total assets and market capitalisation). Across multiple open-source LLMs, including Llama 3 variants and Phi-3, the embedding-based approach performs best particularly for less common statistics, and learning-curve results suggest that only a few dozen labelled observations are often sufficient. The authors also propose a transfer learning approach that can improve estimation for a target variable without labelled data by using the LLM’s text outputs as noisy labels alongside labelled data for other variables, and they report that a reasoning-tuned model does not deliver consistent accuracy gains despite materially higher computational cost. As with other Bank of England Staff Working Papers, the publication is positioned as research in progress intended to elicit comment and debate, and it does not represent Bank of England policy.
Bank of England 2025-10-31
Bank of England staff working paper finds LLM hidden states improve estimation and imputation of economic and financial statistics
The Bank of England's Staff Working Paper No. 1,150 examines using embeddings from open-source large language models (LLMs) to estimate economic and financial variables. A simple linear model trained on embeddings can outperform text outputs for certain statistics. The study covers datasets from the US, UK, EU, and Germany, showing the effectiveness of the embedding-based approach, especially for less common statistics. It also discusses a transfer learning approach, noting that reasoning-tuned models do not consistently improve accuracy despite higher computational costs.