What the model said, against what happened.
Mean absolute error ignores direction. The bias on the same cohort is published on the resolved-fires page, and it is negative: the model predicts shorter than reality.
The newest graded forecasts, most recent at the right. A pair where the orange bar towers over the grey is a fire the model said was nearly over and which burned for weeks.
| Inputs recorded | Forecasts | Fires | Mean error | Median horizon |
|---|---|---|---|---|
| Measured | 231 | 36 | 4.9 d | 5.0 d |
| Partial | 106 | 14 | 5.6 d | 5.8 d |
| Not Recorded | 1,589 | 276 | 21.5 d | 14.4 d |
every graded forecast that stored its inputs, from 27 October. Earlier forecasts are shown as not recorded rather than as modelled -- they may well have had the inputs; we did not keep them. Read the horizon beside each row: these groups differ in how far ahead they were forecasting far more than in what they knew, so the errors are not comparable
| Released | Mean error | Interval | Forecasts | Fires | Measured over | Note |
|---|---|---|---|---|---|---|
| 10 November 2026 | 11.4 d | 10% | 182 | 45 | 6 August 2026 – 15 September 2026 | |
| 7 November 2026 | 11.4 d | 11% | 182 | 45 | 6 August 2026 – 15 September 2026 | |
| 5 November 2026 | 11.4 d | 12% | 182 | 45 | 6 August 2026 – 15 September 2026 | |
| 5 November 2026 | 11.6 d | 12% | 174 | 44 | 6 August 2026 – 15 September 2026 | first navy build; record page published |
How the error is calculated
One forecast per fire per day, scored against the day the fire reported containment.
Mean absolute error is the average gap between predicted and actual days remaining, ignoring direction. A fire forecast on ten days contributes ten tests, so long fires weigh more than short ones.
The interval
8.5% of outcomes landed inside the 80% interval the forecast published.
It should be about 80%. It is not, which means the model is not merely biased but overconfident: the range it offers is far too narrow for how wrong it turns out to be. every graded forecast that stored the spread it published. Not the headline cohort: a forecast could not both end before the 16 September cutoff and have been made after spreads began being stored on 24 October.
What this cannot tell you
These figures describe short-horizon calls.
Nine in ten graded fires ended within about four weeks of the forecast being made, and the error grows steeply with the horizon. The figures say little about a fire with two months left.
Amendment history
Figures are restated when the measurement changes, not when the result is unflattering.
On 27 October the forecast of record moved to after the daily weather fetch, so figures before and after that date are not continuous. Every release is listed above with the figures it carried.