Forecast price
It is calculated from the price seen by the model and its forecast percentage change. Both are stored before the horizon ends.
PUBLIC EVALUATION · PRICE DISTANCE · PROTOCOL V1
A price projection is not well described by a binary label. Here we measure percentage distance from the observed price, both at the expected time and inside an explicit timing margin.
By horizon
Error is the absolute distance between forecast and observed price, divided by the forecast price. Lower is better. The last column compares the model with leaving the initial price unchanged.
| Horizon | Sample | Target time | Best within ±2 h | Best within ±3 h | Within 0.25% in ±3 h | Versus unchanged price |
|---|---|---|---|---|---|---|
| 1min | 76,630 | 0.110% | 0.005% | 0.004% | 96.4% | 0.035 pp worse |
| 10min | 108,748 | 0.196% | 0.017% | 0.014% | 90.3% | 0.024 pp worse |
| 20min | 60,676 | 0.256% | 0.016% | 0.014% | 90.5% | 0.022 pp worse |
| 25min | 49,988 | 0.278% | 0.020% | 0.013% | 90.5% | 0.022 pp worse |
| 35min | 35,571 | 0.324% | 0.033% | 0.016% | 85.7% | 0.020 pp worse |
| 40min | 31,350 | 0.349% | 0.042% | 0.020% | 81.7% | 0.019 pp worse |
| 45min | 28,855 | 0.378% | 0.049% | 0.026% | 78.4% | 0.018 pp worse |
| 50min | 23,703 | 0.363% | 0.052% | 0.026% | 78.3% | 0.018 pp worse |
| 55min | 19,652 | 0.345% | 0.052% | 0.028% | 77.9% | 0.017 pp worse |
| 75min | 14,242 | 0.390% | 0.072% | 0.040% | 72.9% | 0.017 pp worse |
| 65min | 16,328 | 0.366% | 0.061% | 0.033% | 75.9% | 0.017 pp worse |
| 70min | 15,334 | 0.370% | 0.062% | 0.033% | 74.7% | 0.017 pp worse |
| 80min | 13,228 | 0.421% | 0.085% | 0.047% | 71.2% | 0.015 pp worse |
| 85min | 12,298 | 0.448% | 0.093% | 0.051% | 69.9% | 0.016 pp worse |
| 90min | 12,335 | 0.439% | 0.103% | 0.057% | 68.7% | 0.016 pp worse |
| 5min | 182,411 | 0.114% | 0.013% | 0.010% | 93.9% | 0.019 pp worse |
| 15min | 82,614 | 0.213% | 0.014% | 0.012% | 91.5% | 0.020 pp worse |
| 30min | 47,530 | 0.306% | 0.026% | 0.012% | 90.9% | 0.019 pp worse |
| 1h | 20,773 | 0.357% | 0.057% | 0.034% | 75.1% | 0.015 pp worse |
| 2h | 10,475 | 0.474% | 0.136% | 0.085% | 64.2% | 0.013 pp worse |
| 3h | 11,721 | 0.731% | 0.243% | 0.199% | 53.0% | 0.014 pp worse |
| 4h | 8,174 | 0.964% | 0.401% | 0.298% | 47.6% | 0.011 pp worse |
| 6h | 3,035 | 0.875% | 0.475% | 0.380% | 43.7% | 0.010 pp worse |
| 1d | 2,357 | 2.076% | 1.367% | 1.271% | 24.2% | 0.005 pp worse |
The current cohort begins on 10 July 2026. Percentages update as more horizons mature. Earlier rule sets are not blended in.
Method
It is calculated from the price seen by the model and its forecast percentage change. Both are stored before the horizon ends.
We compare with the latest valid close at the target time using the finest available market series. Its timeframe and observation gap are published as well.
We find the observed close nearest to the target before and after the expected time. A window never reaches back before the forecast was created.
We repeat the same calculation as if price had stayed at its initial anchor. This shows whether the forecast added anything beyond natural proximity.
How to read it
That is why two and three hours are separate and the error at the exact target time remains visible.
The JSON also includes distance to candle highs and lows and the time offset of the nearest close.
A 0.5% error may be small for one asset and large for another. Symbol, volatility, sample and horizon should be read together.
It does not include spread, slippage, fees or an entry and exit rule. It evaluates the price projection, not account profit.
The aggregate data is public and can be filtered by symbol or horizon.