Load historical candles
Shares the same browser storage as Candle History & Compare. Import OHLCV history there, then return here. You can also load a CSV directly.
Supports plain .csv and compressed .csv.gz files. Large archives are streamed and the latest 250,000 valid candles are kept in memory to avoid freezing the browser. Recommended: thousands to hundreds of thousands of 1-minute candles covering multiple market regimes. The bundled scorecard and V28 files are prediction records, not OHLCV candles, so they cannot train a candle-feature model by themselves. Full archive source: Bitstamp BTC/USD, 6.85 million rows, Jan 2012–Jan 2025. The model retains at most the latest 250,000 candles in memory; download the raw archive link to keep all rows locally.
Historical predictive test
Chronological 70/30 split; features use only candles available at prediction time. Test labels are never used to fit the model.
| Signal feature | Correlation with next 15m return | Direction | Interpretation |
|---|---|---|---|
| Run the test to calculate feature correlations from loaded candles. | |||
Next 15-minute direction
Uses a logistic model trained on the loaded history. The forecast is timestamped and saved once; it is not continuously rewritten.
Load a valid candle history and run the historical test before generating a forecast.
Probability is a model estimate, not a calibrated real-world chance until calibration is verified. “No edge” is an acceptable model conclusion.
Locked predictions & outcomes
Each locked forecast is append-only. Once the 15-minute horizon has elapsed, score against the candle close if matching candles are available.
| Created UTC | Spot | Forecast | Model probability UP | Horizon end | Actual | Result |
|---|---|---|---|---|---|---|
| No locked predictions yet. | ||||||
How this engine looks for correlations
Returns over 1, 3, 5, 10, 15 and 30 minutes.
Recent realized volatility and candle range relative to price.
Body direction and size relative to the high-low range.
Recent volume versus its trailing average, when volume is available.
Whether the close 15 minutes later is above or below the current close.
Chronological holdout, accuracy, Brier score, log loss, majority baseline and momentum baseline.
Correlation is association, not causation. The current bundled prediction records previously showed only weak relationship between the forecast and outcome, and the always-UP baseline did better in that sample. This new engine is designed to test actual candle-derived features separately rather than assume prior forecasts are informative.