SKILL AND RETURN OF SKILL IN DYNAMIC EXTENDED-RANGE FORECASTS

The skill of a set of extended-range dynamical forecasts made with a m odem numerical forecast model is examined. A forecast is said to be sk illful if it produces a high quality forecast by correctly modeling so me aspects of the dynamics of the real atmosphere; high quality foreca sts may also occur by chance. The dangers of making a conclusion about model skill by verifying a single long-range forecast are pointed out by examples of apparently high ''skill'' verifications between extend ed-range forecasts and observed fields from entirely different years. To avoid these problems, the entire distribution of forecast quality f or a large set of forecasts as a function of lead time is examined. A set of control forecasts that clearly have no skill is presented. The quality distribution for the extended-range forecasts is compared to t he distributions of quality for the no-skill control forecast set. The extended-range forecast quality distributions are found to be essenti ally indistinguishable from those for the no-skill control at leads so mewhat greater than 12 days. A search for individual forecasts with a ''return of skill'' at extended ranges is also made. Although it is po ssible to find individual forecasts that have a return of quality, a c omparison to the no-skill controls demonstrates that these return of s kill forecasts occur only as often as is expected by chance.