The majority of weather derivatives are temperature-based, using degree-days, and cover periods ranging from one week to multiple months. Weather derivatives are usually priced using climatological data, since there is a perceived lack of skill at scales beyond the deterministic weather forecast period of 6 to 10 days. The prevailing view of long-range forecasting was best summed up by a weather trader at the June 2001 Weather for Risk Management Association meeting, who quipped “Seasonal forecasts are garbage.” However, this view is changing due to two recent developments: 1) the demonstration of skillful seasonal forecasts (using non-deterministic methods) operationally and 2) enabling weather traders to construct statistical pricing models by optimally blending forecasts and climatology. This presentation will address both of these topics.

The achievement of forecast skill at seasonal timescales has taken time and significant investment. WSI has been delivering regional seasonal forecasts to its clients since April 2000. The forecast skill in the intervening months was assessed using both mean absolute errors with respect to climatology and directional correctness. In November 2000, WSI started providing specific degree-day forecasts for selected cities on seasonal timescales. From November 2000 to May 2001, WSI validated seasonal forecasts (two per month) covering the following three months each for eleven cities (88 point specific forecasts in all). Eighty three percent of the forecasts (73 of 88) were directionally correct.

The second topic discussed involves the construction of statistical pricing models for use by weather traders using current forecasts and historical weather data. This involves three steps: (1) Developing an understanding of the type of data traders need, (2) building a unique temperature hindcast database to deliver the required information, and (3) developing an operational forecast database to verify the methods established by the hindcasts. The pricing models will require a forecast probability distribution, rather than a precise forecast. This distribution is derived based on both the forecast accuracy and consistency of the hindcasts. Finally, WSI will discuss the development of this unique hindcast data set and will demonstrate conversion of these forecasts into probability space.