European Seasonal Forecast Performance

Claros outperforms detrended climatology in 55-71% of site-season predictions across all variables – and the full probabilistic distribution adds independent value on top of that in the majority of seasons. When Claros flags a strong signal, performance is substantially higher: wind directional accuracy reaches 89-94% at moderate-to-strong signal strength, and the distribution correctly identifies extreme wind seasons with near-certainty when its IQR exceeds 1.6x the historical spread.

“Useful” defined: a site-season prediction is useful if the Claros weighted best estimate is closer to the observed outcome than the detrended climatological mean – a direct test of seasonal skill with the warming trend removed. Distribution useful tests whether the full probabilistic output correctly tilts toward the side where the outcome fell. Both metrics are evaluated against the same detrended baseline throughout.

These three seasons illustrate what acting on a Claros seasonal signal looks like in practice. Each shows the Claros best estimate anomaly (how far Claros positioned its central forecast from the detrended climatological median), what the observed seasonal mean turned out to be, and whether the distribution was also correctly positioned. All figures are averages across German stations in the Study B dataset.

These examples are selected from the Study B (UK/FR/DE) dataset which uses a later forecast issue date than Study A – a commercially realistic lead time. The anomalies are measured against the detrended climatological median for each station. DJF 2017 is the counterexample included in the full dataset: poor for all variables and countries, and shown transparently in the Study B section.

Reading this table: the “vs raw climatology” columns compare Claros against the simple 30-year historical mean – the benchmark most widely used in practice. The “vs detrended” column is the stricter scientific test: it removes the warming trend and asks whether Claros still adds skill beyond what a trend-aware baseline would give. Temperature gains vs detrended drop to ~3%, confirming that much of the raw-climatology improvement reflects warming trend skill, not seasonal anomaly prediction. Wind retains a robust 17% MAE improvement vs detrended, confirming genuine seasonal wind skill. For degree days the detrended baseline MAE is much closer to Claros (84.7 vs 82.3 HDD; 82.0 vs 79.4 CDD), so both comparisons are meaningful. Users who already apply detrending in their own workflows should weight the detrended column.

Signal probability: Claros probability that the outcome falls outside the detrended Q10 (dry/calm tail) or Q90 (wet/windy tail). Observed hit rate: fraction of site-seasons in that bucket where the outcome fell in the relevant tail. Base rate: climatological expectation (~12% below dry threshold, ~25% above wet threshold). Low-signal buckets confirm suppression; high-signal buckets confirm elevation. Both directions are actionable.

What “distribution useful” means: the Claros distribution (Q5–Q95 quantile output) is deemed useful when: (i) the observed outcome was above the detrended climatological median AND Claros’s 5 closest quantile values to the observed outcome were mostly positioned above the corresponding detrended climatological quantiles at the same percentile levels; OR (ii) vice versa on the downside. In plain terms: the distribution is useful when it correctly tilts its probability mass toward the side where the outcome fell, relative to the detrended historical spread. This is primarily a test of directional shift – not of whether Claros correctly predicted the shape or width of outcomes. A site-season can be distribution-useful but not BE-useful (correct directional tilt despite a wrong central estimate), or vice versa – the two metrics are complementary.

Wind is the strongest variable for distribution skill: 71% overall vs 62% for BE – the distribution adds clear independent value across all seasons (DJF 71%, MAM 74%). For temperature, the distribution dramatically outperforms BE in MAM (75% vs 47%). For precipitation, distribution adds a consistent 4–8pp across all seasons.

Metric definition: IQR ratio = (Claros Q75 − Claros Q25) ÷ (Detrended Cli Q75 − Detrended Cli Q25), computed per site per season. Both IQRs are drawn from the same quantile ladder used throughout this study (Claros Q5–Q95 output; detrended climatological quantiles derived from the 1980–2009 baseline for each site). A ratio of 1.0 means Claros is expressing the same spread as the historical distribution. Below 1.0 = Claros is expressing a narrower range of outcomes than historical norms (higher model confidence). Above 1.0 = Claros is expressing a wider range (lower confidence or elevated volatility signal).

What this section is and is not measuring: the y-axis in both charts below is “distribution useful %”, which uses exactly the same definition as the section above – the 5-closest-quantiles test applied locally around the observed outcome. So this section asks: does the width of the Claros IQR, taken as a secondary signal, predict whether the distribution will be directionally useful? It is not asking whether the Claros IQR correctly predicts the actual spread of outcomes. A narrow Claros IQR does not mean outcomes were actually tighter – it means the model expressed more directional confidence, and that confidence correlates with higher distribution-useful rates. The “observed extreme outcomes” chart (right panel) separately asks whether IQR width predicts observed volatility – and for wind this relationship is strong and commercially actionable.

For temperature, a narrow Claros IQR (0.7–0.9×) is associated with 69% distribution-useful vs 61% overall – the model’s confidence expression is a useful secondary signal. For wind, a very wide IQR (>1.6×) is a near-certain volatility alarm: 90.3% of those site-seasons produced an observed extreme (outside detrended Q10–Q90), and 92.5% were distribution-useful. This is the most commercially actionable signal on this page for risk-management users.

Study B uses a later forecast issue date than Study A – a more commercially realistic test. Performance varies meaningfully by country. Germany shows the strongest wind signal; temperature distribution skill is the standout result across all three countries. DJF 2017 is a genuine poor season for all variables and countries – included here for transparency.

The global study validates Claros at both 1-day and 90-day resolution as representative points across a 2-year predictive horizon, with predictions initiated every 5 days. These are validation reference points – Claros generates predictions at any temporal resolution from hourly upward. The 1-day figures are averages across all lead times from near-term through to approximately 730 days ahead, making them a conservative floor on performance at any given lead time.

Wind shows the most striking pattern: 90-day skill is nearly 4× the 1-day figure – far beyond what simple averaging of independent daily errors would produce. This points to Claros capturing persistent sub-seasonal wind anomalies that accumulate through the season. Temperature compounds more modestly in Europe (2.1×), consistent with more frequent intra-season reversals in European temperature regimes. Precipitation skill is similar at both resolutions, reflecting its inherently less persistent seasonal signal in Europe.

The European distribution studies (Study A, Study B) use non-overlapping single-lead-time seasonal forecasts – a stricter test with no overlap benefit. Continuous rolling operational use across consecutive 90-day windows, or at sub-seasonal temporal resolution, would be expected to capture more of this persistent signal structure, particularly for wind.

Wind skill is robust across all regions. Temperature compounds more strongly outside Europe – particularly in regions with larger and more persistent seasonal anomalies – supporting the case that the European figures are conservative. Precipitation shows the most regional variation, reflecting genuine differences in seasonal predictability across climate regimes.

Signal size = weighted best estimate anomaly as % departure from detrended climatological median. Directional accuracy = % where sign of Claros anomaly matches sign of observed anomaly vs detrended median. Expected value = (directional accuracy × average correct anomaly magnitude) − (error rate × average wrong anomaly magnitude), representing average information value per signal. This section focuses on best estimate signal strength and directional accuracy. For distribution shift and IQR spread signals, see the Distribution Width section above.