How three-month seasonal means are built

Three-month means average the monthly forecasts; three-month probabilities are recalculated from member averages. Why the two are not interchangeable.

Why three months

Seasonal outlooks are usually issued for three-month windows because a single month at long lead is noisy while a season averages some of that noise away. Many national meteorological services publish their seasonal outlooks only as three-month means, so the three-month view on this site is the one to compare with those charts.

Means are averaged from monthly anomalies

The three-month ensemble mean anomaly is the equal-weight average of the three monthly ensemble mean anomalies, with the window named by its first month. A September to November window from a September run covers leads 1 to 3. A cell needs a value in every constituent month to get a three-month value, so masked cells stay masked.

Probabilities are recalculated, not averaged

Three-month probabilities cannot be produced by averaging the three monthly probabilities. Instead, each forecast member and each hindcast member is averaged over the three months first, new tercile and quintile thresholds are computed from the averaged hindcast members, and the averaged forecast members are counted against them. This is the C3S method, and it means a three-month probability can be larger or smaller than any of its monthly parts.

Available windows

With six monthly leads there are four three-month windows per run. When comparing runs a month apart, the shared windows are those that both runs can complete, which is three. The period picker shows only those. Cross-year windows are labelled with both years, for example November 2026 to January 2027.

Both monthly and three-month products use the same colour boundaries for a given field, so switching between them does not change the legend. Three-month anomalies are typically smaller because averaging damps the monthly extremes.