How the forecast works
The site forecasts how far Bitcoin is likely to move over the next 1, 4 and 24 hours, not which way. Each range is centred on the current price: over hours, Bitcoin's direction is close to random, and no model tried here beat assuming no change.
How wide the range is
The current model (v2, HAR-RV-J) measures how much Bitcoin has moved over the last 15 minutes, hour, day and week, separating sudden jumps from ordinary choppiness, and combines them in the proportions that predicted best over the past 90 days. It allows for busier and quieter times of day and for weekends, and adds room when a Fed decision or US jobs report falls inside the forecast window. It's refitted every midnight (UTC); if a refit looks wrong, the previous day's version is kept. The width is then calibrated on recent outcomes so the price finishes inside the 90% range about 9 times out of 10.
The previous model (v1) sized the range from a single, fast-reacting average of recent moves. It still runs alongside, scored on its own live record.
Out-of-sample test
Both models recreated their forecasts every 15 minutes over Jun–Oct 2026 (30 May – 27 Oct 2026), each using only the prices available at the time. None of v2's settings were chosen on this period: they were picked on January–May 2026 and checked once here.
| Horizon | 90% range held target 90% |
Average width % of price |
Interval score lower is better |
Change | |||
|---|---|---|---|---|---|---|---|
| v1 | v2 | v1 | v2 | v1 | v2 | ||
| 1 hour | 89.7% | 89.7% ±1.0 | 1.30% | 1.14% | 2.694 | 2.562 | −4.9% |
| 4 hours | 89.8% | 89.9% ±2.0 | 2.44% | 2.32% | 5.294 | 5.216 | −1.5% |
| 24 hours | 88.9% | 88.6% ±4.8 | 7.15% | 6.53% | 15.047 | 14.696 | −2.3% |
Interval score: the width of the 90% and 50% ranges plus a penalty for each miss, in proportion to how far outside the price landed. ± is the 95% chance margin on the hit rate, allowing for overlapping forecasts.
During scheduled releases
In 1-hour windows containing a Fed decision, CPI release or US jobs report, v2's 90% range held 57.7% of the time (v1: 55.8%), against a target of 90%. Releases can still move the price further than the widened range allows.
How much to trust these numbers
- The release-hour figure rests on a small sample: 13 releases in the period, only 3 of them Fed decisions. One or two big surprises move it by several points.
- The 4- and 24-hour gains are small and may be partly noise. A 24-hour window overlaps the next 95 forecasts, so five months hold only about 150 independent 24-hour outcomes (hence the wide ± above). The 1-hour improvement is larger and matched what the model showed on January–May.
- Backtests flatter. The fair test is live forecasts. The site quotes v2's live accuracy once it has 14 days of them; until then it says "collecting live data" and shows these backtest numbers, labelled as such. The live results, misses included, are on the track record page.
What it can't do
No range sees a jump before it starts: the first minutes of a sudden move are missed, and the record shows them as misses. The ranges aren't a prediction of direction, and nothing here is advice to buy or sell.