Evaluating NEX-GDDP-CMIP6 Models at Four Stations in Northern Togo: Seasonal-Cycle Skill, Ranking Robustness and Ensemble Selection

Local evaluation of climate models is essential for reliable climate projections. We evaluated 32 NEX-GDDP-CMIP6 models for monthly precipitation, maximum (Tmax) and minimum (Tmin) temperature at four stations in northern Togo over 1983–2014. One model with unusable temperature data (CMCC-CM2-SR5, Tmin bias +7.8 °C) was excluded; the other 31 were ranked with a Comprehensive Rating Metric (RM) combining bias, root mean square error (RMSE), Pearson’s correlation (R) and Willmott’s index of agreement (d). The seasonal cycle explains 66 to 91% of the observed variance, so these metrics mainly measure it. Median R was 0.87 for Tmax, 0.79 for precipitation and 0.71 for Tmin. At each station, nearly all models had a bias of the same sign, changing between neighbouring stations; its origin is unresolved. Free-running models cannot reproduce observed year-to-year variations, so trends are the main other comparison. Tmin warming in the retained ensemble was faster than observed (+0.38 against +0.22 °C per decade), subject to the homogeneity of the series and to internal variability. Fourteen models had an RM above 0.5. Nine formed a stable core (top 15 in ≥ 90% of resamples), but sub-period cores shared only three of seven models, and neither ensemble consistently outperformed random ones. The selection describes past agreement at four sites. The study offers reusable robustness tests and shows that an unstable ranking requires projections to report the spread across alternative selections.