Background

International Research Institute for Climate Prediction has been issuing global climate forecasts at quarterly intervals in an experimental mode since October 1997. Forecasts for rainfall and temperature are issued for the subsequent 3 months and 6 months. Forecasts are issued for rainfall and temperature in the categories of above normal, near-normal and below-normal. The rainfall cut-offs for each category is based on the wettest, normal and driest 10 episodes for the given season from 1960 to 1990. In addition, it provides warnings regarding extremes. These forecasts are based on global climate simulations by four Global Climate Models which are initialized based on observations of global ocean surface temperatures. These simulations are at a resolution of approximately 250 km and thus Sri Lanka is captured in two grid boxes. Seasonal climate forecasting is a new field and the skill obtained thus far is good for regions such as Indonesia, North-East Brazil, and East Africa. In general, the forecast skill for Asia and Europe is weaker than that for other continents. The regions with the most skill within Asia are South-East Asia and South Asia. However, the skill is likely to improve in the coming years based upon improved observations of Indian Ocean sea surface temperatures, land surface hydrology, Eurasian snow cover and the improvement of GCM performance over South Asia.

Given that the atmosphere is a high dimensional chaotic system, one could not even in the best circumstances, predict precisely the particular trajectory that the atmosphere will take. One can only interpret likelihood's of what the atmosphere may do and this likelihood's may be captured by a probabilistic forecast. The probability forecasts also provide an explicit representation of the uncertainty of the forecasters to the users. However, there are several nuances in relation to how cut-off's are chosen and the impact of extreme weather events that has led some to misinterpret these forecasts.

Both modellers and potential users are quite concerned regarding the quality of forecasts - evaluating forecast quality is a simpler task with deterministic forecasts than probabilistic forecasts. Given that some weight age is afforded for all eventualities, essentially there can be no wrong probabilistic forecasts. However, if the forecasting system consistently affords higher weightage to categories that do not occur, then the forecasting system is less reliable. The reliability of the forecasting system can only be evaluated once there are a large number of forecasts and outcomes. However, the IRI forecasting system has been in place only for a limited time. Hence to provide a simple account of its reliability one is compelled to adopt a rather crude alternative - one based on replacing the probabilistic forecast with a deterministic forecast for its dominant category. Based on this

approach, the track record for the IRI for Sri Lanka of the 3 month rainfall forecasts for the 12 forecasts that has been issued since 1997 is eight hits (observation coincides with dominant category of forecast), three 1-category misses (observation in neighbouring category from dominant category) and no 2-category misses. In JFM 1998 and 2000, IRI did not forecast. The IRI temperature forecasts have a record of 7 hits, 2 1-category miss's and 1 1-category miss's.