Forecast accuracy
Computed 2026-11-14 06:08 UTCHow the gas forecast would have performed, replayed hour by hour over 2026-08-02 to 2026-11-14 (129,781 persisted gas samples). Each decision only sees data from before it — the same inputs the live forecast had — and is scored against what gas actually did. All figures are for the Standard tier.
Last 30 days
- Cheapest window: the predicted cheapest hour of the next 12 landed in the actual cheapest quarter 55% of the time (708 decisions), versus 25% for a random hour and 50% for “yesterday's cheapest hour”.
- “Wait” advice: issued at 587 of 708 decision hours; the picked hour was actually cheaper than sending immediately 52% of the time, with a median saving of 9%.
- Direction: best at 12h, calling 72% of moves of 5% or more correctly versus 51% for always guessing the more common direction.
- Exact gwei: lower average error than “gas stays where it is” at 3h, 12h, and than “same hour yesterday” at 3h. Median error is 74%–81% of the actual price, so treat any single predicted value as a rough range.
Error by horizon — last 30 days
MAE in gwei. The forecast MAE is bold where it beats both baselines. Persistence = “gas stays at the current price”; Yesterday = “same as 24 hours earlier”. Bias is the median of (forecast − actual) / actual — negative means the forecast runs low. Direction counts only moves of 5% or more.
| Horizon | n | Forecast MAE | Persistence MAE | Yesterday MAE | Median error | Median bias | Direction hit | Base rate |
|---|---|---|---|---|---|---|---|---|
| 1h | 719 | 0.565 | 0.397 | 0.557 | 74% | -47% | 52% | 52% |
| 3h | 717 | 0.512 | 0.524 | 0.559 | 77% | -21% | 56% | 51% |
| 6h | 714 | 0.677 | 0.590 | 0.561 | 77% | -4% | 63% | 54% |
| 12h | 708 | 0.570 | 0.602 | 0.564 | 79% | +15% | 72% | 51% |
| 24h | 696 | 0.644 | 0.571 | 0.573 | 81% | +24% | 64% | 50% |
Last 7 days
- Cheapest window: the predicted cheapest hour of the next 12 landed in the actual cheapest quarter 42% of the time (156 decisions), versus 25% for a random hour and 51% for “yesterday's cheapest hour”.
- “Wait” advice: issued at 134 of 156 decision hours; the picked hour was actually cheaper than sending immediately 46% of the time, with a median saving of 0%.
- Direction: best at 12h, calling 61% of moves of 5% or more correctly versus 55% for always guessing the more common direction.
- Exact gwei: lower average error than “gas stays where it is” at no horizon, and than “same hour yesterday” at no horizon. Median error is 79%–90% of the actual price, so treat any single predicted value as a rough range.
| Horizon | n | Forecast MAE | Persistence MAE | Yesterday MAE | Median error | Median bias | Direction hit | Base rate |
|---|---|---|---|---|---|---|---|---|
| 1h | 167 | 0.262 | 0.086 | 0.113 | 90% | -82% | 50% | 50% |
| 3h | 165 | 0.346 | 0.140 | 0.113 | 88% | -73% | 48% | 51% |
| 6h | 162 | 0.574 | 0.165 | 0.114 | 84% | -58% | 48% | 54% |
| 12h | 156 | 0.525 | 0.161 | 0.110 | 82% | -50% | 61% | 55% |
| 24h | 144 | 0.718 | 0.115 | 0.117 | 79% | -33% | 61% | 58% |
Last 90 days
- Cheapest window: the predicted cheapest hour of the next 12 landed in the actual cheapest quarter 62% of the time (2,148 decisions), versus 25% for a random hour and 54% for “yesterday's cheapest hour”.
- “Wait” advice: issued at 1,835 of 2,148 decision hours; the picked hour was actually cheaper than sending immediately 59% of the time, with a median saving of 14%.
- Direction: best at 12h, calling 78% of moves of 5% or more correctly versus 51% for always guessing the more common direction.
- Exact gwei: lower average error than “gas stays where it is” at 6h, 12h, and than “same hour yesterday” at 3h. Median error is 62%–73% of the actual price, so treat any single predicted value as a rough range.
| Horizon | n | Forecast MAE | Persistence MAE | Yesterday MAE | Median error | Median bias | Direction hit | Base rate |
|---|---|---|---|---|---|---|---|---|
| 1h | 2,159 | 0.408 | 0.261 | 0.382 | 63% | -35% | 52% | 51% |
| 3h | 2,157 | 0.373 | 0.358 | 0.383 | 62% | -15% | 61% | 51% |
| 6h | 2,154 | 0.408 | 0.416 | 0.383 | 64% | +4% | 71% | 53% |
| 12h | 2,148 | 0.393 | 0.444 | 0.384 | 68% | +18% | 78% | 51% |
| 24h | 2,136 | 0.422 | 0.383 | 0.386 | 73% | +24% | 68% | 53% |
How this is measured
- Walk-forward replay. At every hour, hour-of-day patterns are re-learned from only the previous 14 days (as the live forecast does), then the forecast is generated from that hour's persisted snapshot. The forecast is deterministic, so this reproduces what the live API would have returned.
- Actuals. The median of persisted samples within ±5 minutes of each target time.
- Cheapest window. A hit means the picked hour's actual price was among the 3 cheapest of the next 12 hours. “Wait” is counted when the picked hour was predicted at least 10% cheaper than now. The live send-now/wait advice is stricter: it also requires the saving to be worth at least $0.05 on a typical swap, so in low-gas periods the site says “send now” far more often than this count suggests. The replay scores the underlying 10% signal because it has no historical ETH price to apply the dollar floor.
- No confidence scores. The forecast does not publish per-hour confidence. Any future confidence label will be defined from these measured rates, not added on top of them.
Model details: methodology. Live forecast: gas fee forecast.
Related ETH gas tools and guides
Gas fee forecastPredicted lower-fee windows over the next 24h–7d.Cheap gas times by dayLowest-fee hours of the week on a 7-day heatmap.Ethereum gas predictionsFull forecast model with 6h, 12h, and 7-day views.ETH gas price todayToday's gas in Gwei plus a 24-hour heatmap.Set a gas alertGet an email when gas drops to your target Gwei.