LLM evaluation in finance: measurement, leakage, and verification

Q: How can LLMs be used as measurement tools in financial and economic research without systematic error?

LLM evaluation in financial research requires accounting for two failure modes before results can be trusted. Prediction bias occurs when LLM outputs are used as regression inputs without correction, producing systematically wrong coefficients. Training leakage occurs when a model recalls rather than reasons after seeing the analyzed historical data. Correction methods should be standard practice for any institution using LLM signals in quantitative workflows.

Q: Why do frontier LLMs disagree on financial sentiment classification, and what does that mean for model risk management?

Frontier and open-weight models show notably lower pairwise agreement on financial transcript sentiment than on standard financial filings. The disagreement is systematic, not random. Model risk management frameworks that rely on a single model's sentiment output without cross-model benchmarking are accepting unquantified variance. A practical way to address this gap is running finance tasks across multiple models and checking agreement scores before any output enters a production workflow.

Q: What did Federal Reserve Bank of New York research reveal about the information markets most need from central bank communications?

A multi-year LLM finance research program at the Federal Reserve Bank of New York found that central bank sentiment carries more predictive weight than quantitative forecasts or hawk-dove signals for markets. The team simulated investor access to Fed deliberation transcripts normally released with a five-year lag and measured how much that information reduced forecast errors. The actual numbers were overshadowed by the tone extracted from the text. For institutions building AI-powered fixed-income analytics, that result reorders the focus of LLM evaluation infrastructure.

Validating LLM outputs for financial and economic research requires strict safeguards like data contamination checks, bias correction, and multi-model cross-verification to keep flawed calculations from corrupting market and risk models.

Mark Vandergon

Lead Data Scientist

Federal Reserve Bank of New York

What you'll take away from this session

Bias correction for AI-generated predictors is solvable

Researchers developed correction methods with software libraries to address the bias LLM outputs introduce when used as regressors in downstream models. Naive two-step approaches produce systematic error; applying these corrections should be standard practice for any quantitative team using AI-derived predictors.

Training leakage is a risk in financial analysis

Holdout tests, memorization probes, and robustness checks using models with earlier knowledge cutoffs can surface training leakage before it contaminates a finding. These tools can detect risks in LLMs that recall rather than reason with trained data from the period being analyzed.

Different models disagree substantially on financial sentiment

Pairwise agreement scores across frontier and open-weight models on Federal Reserve transcript data are notably lower than on standard financial filings. Selecting one model without benchmarking against others introduces a verification gap.

Supervised learning and human annotation remain essential

Fine-tuned smaller models still outperform larger general models on specialized financial classification tasks. Human-annotated validation sets remain necessary to confirm that what the model measures corresponds to what researchers intend.

Reliability lags behind accuracy across agentic financial tasks

Accuracy benchmarks for financial AI tasks are improving rapidly. Reliability measures covering consistency, calibration, and cross-model agreement are not moving at the same pace, a gap that becomes more consequential as agentic workflows extend task horizons from minutes to hours.

LLM finance research has an accuracy and verification problem, yet only one of them is being solved. At the Federal Reserve Bank of New York, Lead Data Scientist Mark Vandergon has spent years building the methodological infrastructure to use language models as scientific instruments in economic research. His conclusion is that LLM evaluation benchmarks for financial tasks are improving rapidly, but the reliability infrastructure needed to trust those results has not kept pace.

Three failure modes structure the session. Using LLM outputs as regression inputs without bias correction produces systematically wrong coefficients. Correction methods exist in published software libraries but are not yet standard practice. Models trained on data covering the period being analyzed may be recalling rather than reasoning, a leakage risk detectable through holdout tests and memorization probes. And when the team ran sentiment classification across frontier and open-weight models on Federal Reserve transcript data, pairwise agreement was substantially lower than on standard financial filings — a finding that applies to model risk management workflows that rely on a single model's output without cross-model validation.

The session's transparency research sharpens the stakes further: when the team modeled market access to Fed deliberation transcripts, reduced forecast error was driven by sentiment in the text, not quantitative forecasts or hawk-dove signals. Institutions building AI-powered analytics around communications should reorder where verification infrastructure matters most.

FAQ

How can LLMs be used as measurement tools in financial and economic research without systematic error?

Why do frontier LLMs disagree on financial sentiment classification, and what does that mean for model risk management?

What did Federal Reserve Bank of New York research reveal about the information markets most need from central bank communications?

Transform the work that matters most

See how Domino helps the world’s most regulated enterprises build, scale, and govern AI-powered applications.

Watch demo Talk to an expert