Projecting the future of rainfall extremes: better classic than trendy

Non-stationarity approaches have been increasingly popular in hydrology, reflecting scientific concerns regarding intensification of the water cycle due to global warming. A considerable share of relevant studies is dominated by the practice of identifying linear trends in data through in-sample analysis. In this work, we reframe the problem of trend identification using the out-of-sample predictive performance of trends as a reference point. We devise a systematic methodological framework in which linear trends are compared to simpler mean models, based on their performance in predicting climatic-scale (30-year) annual rainfall indices, i.e. maxima, totals, wet-day average and probability dry, from long-term daily records. The models are calibrated in two different schemes: block-moving, i.e. fitted on the recent 30 years of data, obtaining the local trend and local mean, and global-moving, i.e. fitted on the whole period known to an observer moving in time, thus obtaining the global trend and global mean. The investigation of empirical records spanning over 150 years suggests that a great degree of variability has been ever present in the rainfall process, leaving small potential for long-term predictability. The local mean model ranks first in terms of average predictive performance, followed by the global mean and the global trend, in decreasing order of performance, while the local trend model ranks last among the models, showing the worst performance overall. Parallel experiments from synthetic timeseries characterized by persistence corroborated this finding, suggesting that future long-term variability of persistent processes is better captured using parsimonious features of the past. In line with the empirical findings, it is shown that, prediction-wise, simple is preferable to trendy.