Multi‐model spread and probabilistic seasonal forecasts in PROVOST

Abstract The skill of the PROVOST (PRediction Of climate Variations On Seasonal to interannual Time‐scales) long‐range multi‐model ensemble integrations is analysed. The ensemble PROVOST forecasts result from integrating three different models over the period 1979–93 using analysed sea surface temperatures. For each model, a set of nine‐member ensembles have been run from consecutive European Centre for Medium‐Range Weather Forecasts re‐analyses. Using the full set of models, a large multi‐model ensemble has been constructed and verified. Positive skill is found for forecasts of geopotential at 500 hPa, temperature at 850 hPa and precipitation. Skilful forecasts tend to occur at the same time in most of the models when skill is computed over large areas; the European region is poorly forecast. The skill commonality may be due to the use of either similar initial conditions or boundary conditions. Skill is shown to be at a maximum in late winter and early spring in mid latitudes. No means have been found for linearly predicting the skill of the ensemble mean using the ensemble spread. The multi‐model ensemble improves the skill of the individual models only marginally when verifying the ensemble mean. However, when using the full ensemble in a probabilistic formulation, the multi‐model approach offers a systematic improvement. The improvement arises both from the use of different models in the ensemble and from the higher ensemble size obtained by combining all of the models for building the multi‐model ensemble. It is shown that a part of the skill improvement in the tropics is due to the multi‐model approach, mainly in spring and summer. On the other hand, most of the gain in the extratropics comes from the increase in ensemble size.