MetSwift – Claros v3 Prediction Performance Validation
MetSwift
Prediction Performance Validation
Europe 2015-2024 – Claros v3 – up to 329 stations

Claros v3 – Prediction Performance Validation

Headline results
Claros v3 against a detrended 1980-2009 climatological baseline
96%
Wind directional accuracy at strongest signal
52% at near-zero signal – Western Europe
90.3%
Seasons producing an observed extreme when wind IQR >1.6x
vs a 3.3% base rate
8.9x
Lift on above-normal wind at strongest signal
89.1% observed vs a nominal 10%
+17.3%
Wind MAE improvement, 90-day
Europe – vs detrended climatology
75%
Spring temperature distribution useful
vs 47% best estimate
89%
Germany wind directional accuracy
12 DJF winters – Study B
Useful = the Claros prediction is closer to the observed outcome than the detrended climatological mean. Distribution useful = the probabilistic output tilts to the side the outcome fell on.
Accuracy rises with signal strength Global Study
Directional accuracy by size of the Claros anomaly – Western Europe – 79 stations – select variable
Prediction in practice Study B – Germany
Claros anomaly vs observed anomaly, against the detrended climatological median

Wind – DJF 2022

51 stations – called above normal
Claros
+1.12 m/s
Observed
+2.04 m/s
IQR
1.24x
92% of sites BE useful – 94% distribution useful

Temperature – DJF 2020

52 stations – record-warm European winter
Claros
+0.26 C
Observed
+1.74 C
IQR
0.95x
100% of sites BE useful – 98% distribution useful

Temperature – DJF 2022

52 stations – mild winter
Claros
+0.79 C
Observed
+0.81 C
IQR
0.89x
100% of sites BE useful – 100% distribution useful
Distribution width as a volatility signal Study A
Claros IQR divided by the detrended climatological IQR – select variable
Distribution useful % by IQR ratio
Observed extreme outcomes by IQR ratio
What the full distribution adds Study A
% of site-seasons beating detrended climatology – 325-329 stations – 39 seasons
Spring temperature: 46.5% best estimate useful, 75.1% distribution useful. Wind gains in every season, reaching 74% in spring.
Tail signal skill Global Study
Observed hit rate against the probability Claros assigned – Western Europe – select threshold, variable and tail
Signal low, outcome suppressed
Signal high, outcome elevated
Low conviction
Nominal climatological rate
UK, France and Germany – DJF winters Study B
157-158 stations – 12 winters 2015-2026 – later issue date – select variable and country
Global reach Global Study
90-day MAE improvement vs detrended climatology by region – select variable
European wind skill compounds from +4.4% at 1-day prediction period to +17.3% at 90-day, a factor of 3.9.
Against ECMWF long-range forecasts Global Study
Claros across a 2-year predictive horizon – vs raw 1980-2009 climatology, the baseline used in published ECMWF verification
Variable, 90-dayClaros MAE impr.Claros seasons beating clim.ECMWF mid-latitude skill at comparable lead
Wind speed+22%~67%SEAS5 European wind verification shows “lack of coherent regions of skill gain”
Temperature+23%~70%Week 3-4 forecasts on average barely beat climatology outside the tropics
Precipitation+22%~54%SEAS5 seasonal skill over European land “low and often indistinguishable from zero”
Sources: Vitart 2025; Johnson et al 2019 (GMD); Crespi et al 2021 (Climate); Haiden and Chevallier 2024 (ECMWF Newsletter 179).
Validated performance, by use