Resolved wildfires

One forecast, made on one day, about one fire. Everything below is counted in these rather than in fires, because a fire forecast on ten days contributes ten tests of the model.

Fires that have left the NIFC active feed, at or above the 100-acre floor the ranking uses. Smaller fires are never forecast, so they can never be graded.

Fires whose end date the fire itself reported and which ended while we were still pulling daily. Only these feed the error and bias below; the rest are listed with their own label in the table.

The typical gap between days predicted and days the fire actually had left, ignoring direction. Half the forecasts were closer than this, half further. Lower is better.

The same gap, keeping its sign. Negative means the model predicted shorter than reality — it says a fire is nearly over when it is not. A model that erred equally in both directions would sit near zero, so this is the figure that says something is systematically wrong rather than merely noisy.

The model never predicts less than half a day. A forecast sitting on that floor can only ever err long, so counting it would drag bias toward zero and flatter the model. Those are excluded from bias and kept in median error, where being wrong is being wrong.

The same bias, for fires that never reported a containment date. Their end is taken from the day they stopped appearing in our pulls — a proxy that runs roughly twelve days long. Shown separately because averaging it with the figure above produces a number about neither.

The share of graded forecasts that landed within two days of what the fire actually did.

Nine in ten graded fires ended within this many days of the forecast being made. The figures describe that range and say little about anything longer — including the 14-to-21 day window a contract runs.