A new AI-powered tool called FedLock ranks Federal Reserve speeches on a hawk-dove spectrum using large language models. Developed by Joe Weisenthal as a hobby project, the tool analyzes thousands of FOMC member speeches from the past 30 years, removing speaker names to eliminate bias and comparing pairs to determine relative hawkishness.

FedLock builds on Weisenthal's earlier Havelock project, which measures how written text compares to spoken language using Walter Ong's orality-literacy framework. The system works by having an LLM judge pairs of speeches side-by-side, considering economic conditions at the time, then converts the results to a 1-100 scale. The tool costs about a penny per new speech analyzed and could potentially be adapted to rank earnings calls on management confidence or pricing power.

The project demonstrates how LLMs can create novel analytical frameworks for financial markets. While Weisenthal acknowledges there's no way to prove perfect accuracy, the resulting time series tracks reasonably well with metrics like the Taylor Rule. The tool represents an innovative application of AI to central bank analysis, potentially offering investors new ways to gauge monetary policy sentiment shifts.