Evaluating TIGGE Rainfall Forecasts for Tropical Eastern Africa

This study evaluates the skill of ensemble rainfall forecasts at lead times up to 9 days for tropical eastern Africa, based on the Thorpex Interactive Grand Global Ensemble (TIGGE). We examine the biases and skill of four Ensemble Prediction Systems (EPS) from Canada (CMC), Europe (ECMWF), the UK (UKMET) and USA (NCEP) as well as in a multi-model average, using a broad array of metrics and verification tools. We find that by most metrics, ECMWF is the most skillful of the individual models at predicting fluctuations in daily rainfall, while CMC shows the most realistic ensemble spread and thus has the best reliability in terms of estimating forecast probabilities, including those of extreme events. The skill of a multi-model average of the four models surpasses that of any individual model, and is skillful for lead times up to at least 9 days . A quantile-to-quantile (Q2Q) mapping bias-correction of the forecasts not only effectively removes systemic biases, but also improves the RMSE skill of all models and the multi-model average. The improvements are greatest when the Q2Q mapping is done separately for each month, highlighting the importance of seasonal dependencies in the biases. Forecasting the most extreme events (>90th percentile) is a challenge for this region, with false alarm ratios just under 50% even for optimized next-day 24 hour forecasts. Verification at larger spatial and longer temporal aggregation scales improves these results, suggesting that the utility of the forecasts will depend on the required forecast resolution (e.g., the averaging period, or the size of the river catchment if used for streamflow forecasting). Temporal and spatial structures in the rainfall data are diagnosed and results are discussed in the context of the unique challenges of forecasting convection in the tropics.