Frequently asked questions

It runs a simulation of an FOMC meeting. It populates a virtual committee with language-model "personas" of the real participants, gives them the economic information available before the meeting, lets them deliberate and vote, and reports the rate decision they reach — plus a statement, minutes, and dot plot. More →

It produces a forward-looking simulation of the decision, which is not the same as a statistical forecast. It is a model of how the committee reasons, not a crystal ball and not a feed from inside the Fed. Useful for intuition; not something to trade on.

No. It is not affiliated with or endorsed by the Federal Reserve. The personas are models of real people built from public records, and they do not see the confidential staff materials real members read. More →

On a backtest of 17 meetings (Mar 2024 – Mar 2026) it called the direction correctly 94.1% of the time (16/17), with a mean error of 1.5 basis points and 82.8% vote alignment. Each meeting was simulated using only pre-meeting data and scored blind. Full table →

No. That figure is a backtest on a small sample of mostly-calm meetings, from one model configuration. It is best read as "this approach has been promising on recent history," not as a probability for your run. Your own run, with your own provider and model, is a different experiment.

Not necessarily. Each run is one random draw of a stochastic process — the debate, the vote split, and sometimes the call itself can change between runs. The research behind this engine runs it ~100 times to get a distribution; the public demo currently runs it once. Treat a single run as one opinion, not a settled estimate. More →

At turning points and during novel shocks with no historical analog — regime changes, surprise data revisions, political-pressure episodes. Its single backtest miss (7 May 2025: it called a cut, the Fed held) was exactly such a turning-point meeting. More →

Never. For any meeting whose outcome has not been published, the "Committee's Call" card stays blurred — we would rather show nothing than invent a number. More →

A packet of current macro indicators (from FRED/BLS), a regional economic narrative, recent news and member speeches, and the prior meeting's decision and statement — all dated before the meeting being simulated. More →

It is a synthetic version: the most recent official Beige Book combined with current web research across the twelve districts, to keep the regional picture up to date between the Fed's eight-times-a-year releases. It is a sourced approximation, not an official document. More →

Every simulation is built only from data dated before the meeting, and the real decision is served on a separate channel the simulation never reads. That point-in-time discipline is what makes the backtest a genuine out-of-sample test. More →

It stays in your browser tab's memory and the server's memory only for the length of one run, travels over an encrypted connection, is never stored or logged, and is scrubbed from any error messages. More →

A run is roughly ~28 model calls against your own API key — about $0.20 on the reference configuration; your provider and model choice will change that. The engine caps spend by stopping abandoned runs and enforcing a hard time limit. More →

You bring a key for a supported provider (Gemini, OpenAI, or Anthropic). Note that the published accuracy figures are for one specific configuration, so a different model is a different experiment and may perform differently.

It is a research and educational simulation. It does not price assets, does not know what markets have already priced in, and is not investment, financial, legal, or tax advice. Do not make financial decisions based on its output.