Waiting for price
Live public market data
Are the predictions useful?
Measured after the outcome arrives. Lower error is better.
Spaced samples avoid overlapping target windows. They can still be correlated.
How this experiment works
A small online model estimates direction and movement from recent returns, volatility and traded-volume imbalance. It starts at 50/50 and learns only after recorded outcomes arrive. These probabilities are experimental, not calibrated confidence. Direction probability and average movement are separate estimates and can disagree. Up probability compares up versus down outcomes, excluding unchanged prices. The blue markers show independently estimated 10-second, 30-second and 60-second endpoint prices from the same saved issue time, up to the selected horizon. Each marker shares the actual price and time axes. The target strip gives the predicted price and change from the issue price. These endpoint estimates do not define the intervening Bitcoin price path. Missing or mismatched issue-time estimates are shown as unavailable. The price display checks the latest received trade four times per second; unchanged trade prices remain unchanged. Enable Compare history to see older forecast endpoints.
A forecast is issued every second for each horizon after a 60-second warmup. All forecasts are saved; only spaced samples train the model. A flat price is reported separately and excluded from direction scores. A 50/50 prediction receives half credit for direction accuracy.
Outcomes use the first sampled trade at or after the target, no more than 3 seconds late. Stale feeds pause forecasts; missed deadlines stay in the ledger as missed. The chart shows one-second observations. A separate raw archive keeps every received aggregate-trade, best bid/ask and top-20 depth message, including rejected inputs and connection events. Book data is being collected for future models; it is not yet a model input. Movement is shown in basis points: 1 bp = 0.01%.
The baseline probability uses only previously resolved spaced samples. Movement error is compared with predicting no price change. No orders, wallet, leverage or profit simulation is connected. Direction accuracy is not trading profitability, and Binance prices are not Polymarket settlement prices.
Prediction ledger
Every forecast saved before its result. Select a row to inspect its price path.
| Issued | Forecast | Up probability | Predicted move | Actual move | Result | Sample |
|---|