Abstract
A strong dependency of model performance in quantitative precipitation forecasts (QPFs) as measured by scores such as the threat score (TS) on rainfall amount (i.e., the better the model performs when there is more rain), is demonstrated through real-time forecasts by the 2.5-km Cloud-Resolving Storm Simulator (CReSS) for 15 typhoons in Taiwan in 2010–12. Implied simply from the positive correlation between rain-area sizes and scores, this dependency is expected to exist in all regions, models, and rainfall regimes, while for typhoon QPFs in Taiwan it is also attributed to the model’s capability to properly handle (within 72 h) the processes leading to more rain, which are largely controlled by the typhoon’s track, size, structure, and environment, and the island’s topography. Because of this dependency, the performance of model QPFs for extreme events can be assessed accurately only when forecasts targeted for periods of comparable rainfall magnitude are included for averaging. For the most-rainy 24 h of the top-5 typhoons, the 0–24-h QPFs by CReSS have mean TS of 0.67, 0.67, 0.58, 0.51, and 0.32 at thresholds of 25, 50, 130, 200, and 350 mm, and 0.64, 0.57, 0.37, 0.33, and 0.22 from 48–72-h QPFs, respectively, suggesting superior performance even 2–2.5 days in advance. These scores are strikingly high, precisely because Taiwan can receive extreme rainfall from typhoons. For smaller (nonhazardous) events, the mean scores are progressively lower, but also unimportant and less representative statistically. Therefore, it is inappropriate to average scores over multiple forecasts as those for less-rainy periods would contaminate the result for key periods. The implication for forecasters is also discussed.
