A Financial Brain Scan of the LLM
This provides a transparent tool for social science researchers to analyze and correct biases in LLM-based economic predictions.
The paper tackled the problem of understanding and steering LLM-generated economic forecasts by mapping them to plain-English concepts like sentiment and technical analysis, and showed that models can be adjusted for risk-aversion or optimism without performance loss.
Emerging techniques in computer science make it possible to "brain scan" large language models (LLMs), identify the plain-English concepts that guide their reasoning, and steer them while holding other factors constant. We show that this approach can map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance. We also show that models can be steered to be more or less risk-averse, optimistic, or pessimistic, which allows researchers to correct or simulate biases. The method is transparent, lightweight, and replicable for empirical research in the social sciences.