Overall, the IRI seasonal climate forecasts outperform the atmospheric general circulation models and ENSO-based empirical predictions on which they are based.
Predictions of seasonal climate anomalies have been made in some form for centuries. One of the earliest scientifically based schemes is that of India's Meteorological Department, which has been issuing predictions for the all-India monsoon rainfall since the late nineteenth century using various statistical methods (Blanford 1884; Walker 1923;Bhalme et al. 1986). One of the more recent developments is the regular production of seasonal climate forecasts that are based, at least in part, on global dynamical climate models [e.g., the U.K. Met Office (UKMO) . These forecasts are the subjective assessment and consolidation of many climate prediction tools, and are thus called "net assessments." They provided probabilistic forecasts for below-, near-, and abovenormal precipitation and temperature for the upcoming 3-month period and the subsequent 3-month period (more information available online at http:// irj.columbia.edu/climate/forecast/net__asmt/). Since mid-2001, however, the IRI has issued these forecasts monthly for the upcoming four overlapping seasons.sinceValidation studies of the prediction tools used by various meteorological centers suggest that they can provide potentially useful information on seasonal climate variability for many parts of the world. For many regions physical dynamical models replicate, and in some cases improve upon, the skill realized from empirical statistical models, and this has given greater credibility to seasonal predictions obtained by both approaches. Thus, regional managers in sectors such as agriculture, energy, and water resources have a 25%-40% probability of extremely above-normal rainfall, which we define as an approximate doubling of risk over the climatologically expected 15% probability. This validation study concerns only the threecategory forecasts.Although the majority of the constituent prediction tools for the IRI net assessments are objective, the 1 Reliability refers to the correspondence between the probabilities given in the forecasts and the subsequent relative frequencies of occurrence in the observations. For reliable forecasts, these two quantities are equal (see Wilks 1995;Murphy 1993Murphy , 1997.2 Where no region is explicitly designated, a forecast for climatological probabilities is implied. For precipitation, explicit forecasts are also not issued for "dry regions," which are indicated. Over these areas the climatological rainfall is so low (less than 30 mm over the 3-month season) that the below-normal category cannot be well distinguished from the near-normal category.
| BAI1S-DECEMBER 2003subjective element of human intervention is not negligible as they are combined to arrive at the final forecast product. Recent improvements to the IRI forecast system have greatly reduced the amount of subjective work necessary to produce the net assessments (Barnston et al. 2003). More objective appr...