Filings, earnings calls, and newswire carry signal that most factor models never use. The obvious way to use them is to score the full history with a current-day language model. But a current-day model knows what came after each document, and that knowledge leaks into every score. The signal looks strong in the backtest for the wrong reason, then fades in production.

ChronoLLM scores each document with the vintage matching its date: a separate version of the model, trained only on text available up to that point. A sentiment score on a 2009 filing reflects what a reader could have known in 2009. The output is a point-in-time series, one value per document, with every row carrying the vintage that produced it. It is delivered as Parquet or CSV and joins directly into factor and risk models.

Any text task becomes a point-in-time series

The tasks are defined by your prompts. Any classification, rating, or extraction you can state in a sentence can be run across the full document history: management tone on earnings calls, event tags on headlines, risk language in filings. ChronoLLM matches each document to its vintage automatically. The rest of the pipeline is your existing data engineering, unchanged.