On the interpretation of seasonal Southern Africa precipitation prediction skill estimates during Austral summer

Differences between two types of prediction skill estimates over Southern Africa are illustrated to better inform the users of seasonal precipitation forecasts over the region who desire assessments of forecast accuracy. Both seasonal precipitation prediction skill estimates for the African continent south of 15°S during the December–March rainy season are derived from the perfect-model method. The perfect-model method is based on a 40-member ensemble of Community Atmosphere Model version 5 simulations forced by observed time-evolving boundary conditions during 1920–2016. The first skill estimate is based on the verification of an ensemble mean forecast spanning many seasons and therefore unconditional on a single boundary forcing. The second skill estimate is based on the verification of an ensemble mean forecast for a single season and is therefore conditional on that year’s boundary forcing. Unconditional prediction skill calculated in 30-year increments for each of the 40 possible forecasts reveals: (1) large spread in skill among the individual forecasts for any given year and (2) temporal variations in skill for each forecast. The magnitude of conditional prediction skill varies greatly from 1 year to the next, revealing that the boundary conditions offer little prediction skill during some years and comparably large skill during others. The simultaneous behaviors of the El Niño–Southern Oscillation and the subtropical Indian Ocean Dipole are related to the largest conditional precipitation prediction skill years. Unconditional skill estimates may therefore mislead users of forecasts who desire assessments of forecast accuracy. Unconditional skill may be temporally unstable, and unlike conditional skill, is not representative of the skill for a given season.