Domain corpus
Federal Reserve, BLS, and Census narratives anchor the model in official economic language.
02 / EconBERT
Economic direction,
not generic emotion.
General sentiment models often confuse cautious wording with bad economic news. EconBERT instead asks what a passage implies for macroeconomic conditions—so easing inflation pressure can be positive even when the prose sounds restrained.
Try the interactive examplesLLM-assisted labels
Pre-2013 passages receive positive, neutral, or negative macro-direction labels.
Distilled classifier
A compact DistilBERT student learns the task, then remains frozen out of sample.
Real-time index
Passage scores become source-balanced monthly measures suitable for forecast evaluation.
Why domain-specific tone?
Language becomes a measurable economic signal.
EconBERT is designed around economic interpretation rather than everyday affect. It separates what a document sounds like from what its information implies for growth, inflation, labor markets, or financial conditions.
The resulting passage-level probabilities are aggregated into monthly narrative indices and evaluated in genuinely real-time forecasting settings, alongside macroeconomic data vintages and professional forecasts.
Research visualization
Narrative indices across the business cycle.
The figure keeps the model comparison and topic-conditioned measures while omitting the alternative-monthly-weightings panel.
Research interests
Domain-specific language modelsText-as-data measurementReal-time forecast evaluationHuman and model expectations