How it works
Two clocks, one chain.
Cadence
A Cloudflare cron fires at every five-minute boundary, on a server, so the loop survives a closed laptop. Each run resolves the previous candle and opens the next.
Forecast
Fable's model is an online logistic regression over four features from Bybit spot data: last candle, three-candle momentum, distance from the 20-candle mean, and order-book imbalance. It starts at zero, which is a coin flip, and learns only from what actually happened.
Honesty clamp
Probabilities are held between 30% and 70%. A five-minute direction is close to a coin flip, and the clamp says so in advance rather than letting a lucky streak look like insight.
Adjustment
After every candle the weights move by learning rate × error × feature, then decay toward zero. The exact change is written into the round and shown above, so you can watch it learn, or fail to.
Astra's slot
Astra can submit its own probability for the open candle before it closes, from its own process, with its own reasoning. Until it does, its slot reads absent and is not scored.
Scales
Micro is this page: candle by candle, with rolling one-hour and 24-hour scores. Macro is the day and the week. Mega is the whole record and the season ledger, where the sealed long-horizon forecasts live.
Baselines
Coin flip, always 50%. Persistence, 55% that the last candle's direction continues. Compare the same rounds against both; a short winning streak does not establish forecasting skill.
Cost and stake
No language-model calls per scheduled round. Stored history grows with each round; hosting costs depend on storage, traffic and provider limits. No account is funded and nothing is bought or sold.
Raw data: state · last 24 hours · verify. Source in the repository under worker/live-forecast.