marketprice: the loop, the scoring, and what a good result would look like

Lab notes for marketprice.org: how the state is built, why Brier score, and the bar the model has to clear.

Every minute, just after the candle closes: fetch sixty one-minute candles of BTC-USD from Coinbase's public feed; build a state of numbers only — the last twenty closes, percentage changes over 1, 5, 15 and 60 minutes, RSI(14), the last five volumes, the fifteen-minute range; ask Jev whether the next close is higher (choice), how strong the trend is (score 0–3) and whether five-minute momentum is positive (noul); store the answers; settle every earlier prediction whose next candle is now known.

scoring

Accuracy alone rewards a model that says "up" whenever the base rate is above half. The Brier score — the mean squared distance between the probability given and the outcome — punishes confident wrong answers and rewards honest uncertainty. A coin flip scores 0.25. Persistence (the next minute goes the way the last one went) and "always up" are scored on the same minutes, so the model is judged against the cheapest possible rivals.

the bar

A Brier score below 0.25 over a thousand settled minutes would mean the probabilities carry information.
Beating persistence on accuracy over the same span would mean the information is not just momentum.
Neither is expected. If both happen, the next question is whether it survives fees, and the answer to that is usually no.

Live page: marketprice.org. Code: lab/marketprice.