marketprice lab

Can a System One model call the next minute of a price? A live experiment, scored every minute against what actually happened.

marketprice.org asks TypeSafe's Jev one typed question every minute — will BTC-USD's next one-minute close be higher? — hands it sixty minutes of candles as numbers, stores the probability it returns, and one minute later stores the real close next to it. The page shows the model's accuracy and Brier score beside two baselines (persistence and "always up") and the coin flip.

A calibrated classifier is not a forecaster, and minute-level price moves are mostly noise. The honest expectation is that nothing beats the coin flip by much. The experiment exists to measure that instead of assuming it, and to run our own smaller model on the same loop once there are enough labelled minutes.

what it costs

One Jev call a minute is about $0.00002; a day of the experiment costs three cents. The page is static, rebuilt every minute, with no scripts and no cookies.

what comes next

More products (ETH-USD, a stock index during market hours) and more horizons (5 and 15 minutes).
The same loop with decide, our local model, trained on the minutes Jev has already labelled.
A published dataset of every prediction and outcome.

marketprice.org · lab notes