Fig. 2: Skill relative to linear regression.
From: Standard assessments of climate forecast skill can be misleading
Dynamical model Niño3.4 random walk skill score (RWSS) at lead 3 months relative to linear regression model using the methods: biased, unfair, fair, fair-sliding, and fair-all in panels (a–e). Forecasts are for the period 1999 to 2016. The white shaded area is an envelope encompassing the RWSS that would be obtained 95% of the time for coin toss trials. The blue shaded area indicates regions where a given dynamical model would be closer to observations than the regression model more often than expected due to chance. The tan region is the converse where the regression model is more often closer to observations than the dynamical model than expected by chance. The RWSS can vary between −1 (model always less skillful than regression), 0 (model and regression equally skillful), to 1 (model always more skillful than regression). This score can also be expressed as the percentage of wins for the dynamical model (right vertical axis). The panels (f–j) show the final value of the random walk skill score, RWSSn, for lead times from 1–11 months. The coloured boxes in panels (f–j) are ordered to show the model rankings relative to linear regression at lead 6, going from least skillful on the left to most skillful on the right. The colour code for the models is given in the legend in panel a. For skill relative to logistic regression, see Supplementary Note 3.
