Bank of England Researchers Reveal Hidden Economic Knowledge Inside AI Models

Artificial intelligence has long promised to revolutionize data analysis, but a new study from the Bank of England suggests that large language models (LLMs) may hold more economic insight than they reveal through words alone. The research, authored by Marcus Buckmann, Quynh Anh Nguyen, and Ed Hill, explores how the hidden layers of LLMs can be used to estimate and even impute economic and financial statistics with remarkable accuracy often outperforming the text-based responses these models provide directly.

The Hidden Value in Hidden States

LLMs like Llama 3 or Phi-3 are trained on vast swaths of text, including data that implicitly encodes economic indicators. While these models may not always provide accurate textual answers to economic questions, the researchers demonstrate that the internal hidden states also called embeddings contain far richer and more structured information. By training a simple linear regression model on these embeddings, the team found that they could predict key variables such as unemployment rates or firm-level financials more precisely than by analyzing the models' textual outputs.

This discovery is significant for investors and analysts alike. It implies that LLMs inherently represent a latent map of the economic world one that can be mathematically decoded even when it's not explicitly articulated. In other words, the information is there, but hidden beneath the surface of what the AI 'says.'

Unlocking Economic Intelligence Through Embeddings

The study focused on datasets covering regional and corporate statistics across the United States, the United Kingdom, the European Union, and Germany, as well as firm-level financial data for U.S.-listed companies. The researchers applied a ridge regression model referred to as the Linear Model on Embeddings (LME) to estimate variables such as GDP per capita, unemployment rates, and firm profitability. The LME consistently outperformed direct text-based predictions, especially for less common or granular economic indicators.

For example, while an LLM might fail to accurately state the percentage of mortgages in arrears in a particular county, its embeddings encoded sufficient contextual information to infer that figure more reliably. These results held true across multiple model families, including smaller and quantized versions of the Llama 3 and Phi-3 models, demonstrating the robustness and scalability of the approach.

Small Data, Big Insights

One of the study's most intriguing findings is that the LME requires only a few dozen labeled examples to achieve high accuracy. This feature drastically reduces the cost and effort typically required for supervised learning, making it especially appealing in finance and macroeconomic research where labeled data is scarce or delayed.

The team's learning curve analysis showed that with as few as 25 data points, the LME already outperformed direct text responses. This opens the door to rapid and cost-efficient applications such as forecasting regional economic health, estimating missing firm data, and detecting anomalies in financial datasets all using existing open-source LLMs.

Beyond Text: Efficiency and Transferability

The researchers also compared their method to advanced reasoning LLMs models that engage in multi-step thinking before generating responses. Surprisingly, even these reasoning models did not match the accuracy of the LME. Despite producing longer and more reflective answers, their text-based outputs remained less consistent than the embeddings-driven estimates. Moreover, the LME approach proved to be orders of magnitude more computationally efficient, making it more practical for large-scale or real-time applications.

Further experimentation revealed that the embeddings could be effectively used for transfer learning. Even when the LME lacked any labeled data for a specific variable, the model could leverage embeddings trained on related variables to produce reasonable estimates. While this transfer learning method did not always surpass text outputs, it provided a promising pathway for zero-shot estimation in scenarios with limited data availability.

Applications for Data Imputation and Super-Resolution

Two of the most practical applications highlighted in the study are data imputation and super-resolution. In financial and economic datasets, missing or incomplete values are a common challenge. The research demonstrated that embeddings can be integrated into imputation models to fill these gaps more accurately, improving the quality of statistical analyses and risk models.

Similarly, for super-resolution tasks such as estimating economic indicators at sub-regional levels from national data the embeddings offered a novel way to infer fine-grained insights from broader aggregates. This could allow policymakers and investors to assess local economic health with a precision previously reserved for official statistical agencies.

Implications for Investors and Data Practitioners

The implications of this research extend far beyond academic curiosity. For investors, fund managers, and data scientists, the ability to extract structured economic intelligence from LLM embeddings could redefine how alternative data is sourced and analyzed. Instead of purchasing pre-cleaned datasets or waiting for delayed releases from government sources, analysts could generate near-real-time estimates from public open-source models.

Moreover, the findings suggest that as LLMs continue to scale, their internal representations of the world will only grow more accurate and comprehensive. Larger models encode spatial, temporal, and financial relationships with increasing precision, as shown by the team's observation that performance improved steadily with model size from 1 billion to 70 billion parameters.

For institutions that handle large-scale data processing, such as central banks, investment research firms, or asset managers, embedding-based modeling could streamline operations, reduce dependence on external vendors, and open up new analytical frontiers in economic forecasting and portfolio risk assessment.

Looking Ahead

While the Bank of England researchers emphasize that the views expressed in their paper do not represent official policy positions, their findings hint at a profound shift in how artificial intelligence may intersect with economic measurement. If hidden layers of LLMs already capture the essence of regional and corporate dynamics, then the next generation of financial analytics may rely less on what models can say and more on what they quietly know.