Estimate forecast reliability by lead time.
Forecast skill decays roughly exponentially with lead time as small initial errors amplify — the chaotic behaviour that limits deterministic prediction to about a fortnight. Changeable regional climates shorten that horizon further. Ensemble agreement is the second factor: when models converge, confidence at a given lead time is much higher than when they diverge, which is why agreement scales the raw skill rather than replacing it.
Forecast confidence decay
Skill = 100 x exp(-lead days x volatility / 7); confidence = skill x (0.5 + model agreement / 200)
One to three days is generally reliable, five days is a useful trend, and beyond about ten days deterministic forecasts approach the skill of climatological averages.
Models start from slightly different analyses of an incompletely observed atmosphere, and small differences amplify. Where they agree, confidence is genuinely higher.
Many model runs with slightly perturbed starting conditions. The spread of outcomes is itself the uncertainty estimate, which is what the agreement input represents here.