Seasonal prediction of Horn of Africa long rains using machine learning: The pitfalls of preselecting correlated predictors

Abstract

The Horn of Africa is highly vulnerable to droughts and floods, and reliable long-term forecasting is a key part of building resilience. However, the prediction of the “long rains” season (March–May) is particularly challenging for dynamical climate prediction models. Meanwhile, the potential for machine learning to improve seasonal precipitation forecasts in the region has yet to be uncovered. Here, we implement and evaluate four data-driven models for prediction of long rains rainfall: ridge and lasso linear regressions, random forests and a single-layer neural network. Predictors are based on SSTs, zonal winds, land state, and climate indices, and the target variables are precipitation totals for each separate month (March, April, and May) in the Horn of Africa drylands, with separate predictions made for lead-times of 1–3 months. Results reveal a tendency for overfitting when predictors are preselected based on correlations to the target variable over the entire historical period, a frequent practice in machine learning-based seasonal forecasting. Using this conventional approach, the data-driven methods—and particularly the lasso and ridge regressions—often outperform dynamical seasonal hindcasts. However, when the selection of predictors is done independently of both the train and test data, by performing this predictor selection within the cross-validation loop, the performance of all four data-driven models is poorer than that of the dynamical hindcasts. These findings should not discourage future applications of machine learning for rainfall forecasting in the region. Yet, they should be seen as a note of caution to prevent optimistically biased results that are not indicative of the true power in operational forecast systems.