A point‑in‑time archive for forecasting research
Chronology records news and public data with timestamps, so AI agents can be trained and tested on the past without seeing the future.
Will the central bank cut rates at its next meeting?
t = Sep 14the problem
Forecasting agents are trained and tested on questions whose answers are already known. A fair setup gives the agent only what was known at the time. Most data sources return today's version of the world instead: revised statistics, renamed pages and articles published after the question closed. A backtest built on them leaks the answer.
Chronology answers each request as of a time t and returns only what was public by then.
data
| data | source | as of t |
|---|---|---|
| news | — | Articles published by t |
| economic data | FRED | Values as first reported by t |
| wikipedia | — | The revision at t |
| web pages | Internet Archive | The latest snapshot by t |
| prediction markets | Kalshi, Polymarket | Prices at t, outcomes settled by t |
research directions
forecasting
Agents that research a question and forecast it, trained and tested with backtests that can't leak.
agent systems
Point-in-time sandboxes, where the tools an agent calls (search, web, data) answer as of t, for training and evaluation at scale.
retrieval
Search that finds what an agent needs at a given moment, measured on real agent queries.
data quality
Clean, dated events and entities from raw news, with automated checks and repairs that keep the archive accurate.
continual learning
Models that keep learning as time moves forward, trained only on what was known at each moment.