Towards the prediction of the East Africa short rains based on sea‐surface temperature–atmosphere coupling

It is shown that the July–September sea-surface temperature (SST) pattern contains moderately strong relationships with the October–December (OND) seasonal rainfall total averaged across East Africa 15°S–5°N, 30°–41.25°E. The relations can be described by using three rotated global SST empirical orthogonal functions (EOFs), mainly measuring aspects of SST patterns in the tropical Pacific (related to El Niño/Southern Oscillation), tropical Indian and, to a lesser extent, tropical Atlantic. Confidence in the relationships is raised because the three EOFs correlate significantly with OND near-surface divergence over the tropical Pacific, Indian and Atlantic Oceans (extending into Northern mid-latitudes), as well as with the rainfall in East Africa and also with rainfall across southern and western tropical Africa. For the East African region, multiple linear regression (MLR) and linear discriminant analysis prediction models are tested. The predictors are pre-rainfall season values of the three rotated SST EOFs. The predictors use information through September. Validating MLR hindcasts using a 1945–1966 (1967–1988) training period and a 1967–1988 (1945–1966) testing period between 30 to 60% of the area-averaged rainfall variance is explained. To achieve unbiased estimates of the expected skill of a forecast system, it is safest to keep model training and testing periods completely separate. The above strategy achieves this in the most important step of ensuring that the models fit the SST predictors to the rainfall predictand using years independent of the testing period. However, the EOFs were calculated over 1901–1980, so for hindcasts prior to 1981, the EOFs describe the SST variability a little better than could be achieved in real-time, which could inflate skill estimates. Tests in the years 1981–1994, independent of the 1901–1980 eigenvector analysis period, do produce similar levels of skill, but a few more forecast years are needed to confirm this result. It is shown that the mean verification at each individual location within East Africa is somewhat lower, which is important to consider for some applications. The need to monitor the prediction relationships and update the models is emphasised. Furthermore, these forecasts only become available as the OND season is underway, though some evidence is found for one of the EOF predictors having skill as early as June. © 1998 Royal Meteorological Society