Why seasonal anomalies use a 1993–2016 hindcast climatology

Seasonal forecast anomalies are measured against each model’s own hindcast climate, not observations. Why the reference period matters when comparing models.

Model climate, not observed climate

Every anomaly on this site is measured against the forecasting system’s own hindcast climatology, not against observations. A hindcast, or reforecast, is the same model run for past years as if it were forecasting them. Averaging those runs gives the model’s expected value for each start month, lead and location. Subtracting it removes the model’s systematic bias and leaves the forecast signal.

Why 1993 to 2016

Copernicus requires every contributing centre to provide hindcasts for at least 1993 to 2016. Using the same 24-year window for all systems is what makes their anomalies comparable: a +1 °C anomaly from ECMWF and a +1 °C anomaly from DWD are both departures from the same era. Some centres run longer hindcasts, but only the common period is used here.

Consequences for reading the maps

Because the reference period ends in 2016 and the climate has warmed since, a temperature anomaly of zero is cooler than a recent-years average would suggest, and widespread positive anomalies are the expected background rather than a striking signal. This is a property of every C3S seasonal product, not a quirk of this site. Compare the strength of anomalies between regions and models rather than reading them as departures from what recent years felt like.

The reference period also matters for probabilities: tercile and quintile thresholds are the empirical percentiles of hindcast members pooled across the same 24 years, computed separately for each start month and lead. A different reference period would move every threshold and change every probability.

What is not corrected

The anomaly removes the mean bias but not errors in variability or in the shape of the distribution. If a model’s hindcast winters are too calm, its forecast anomalies will be too small even after subtracting the mean. The reference period and member counts used for each layer are recorded in the site’s provenance data.