Seasonal Forecasting of Tropical Cyclones, Tropical Rainfall and the El Nino-Southern Oscillation

Dr. Christopher W. Landsea

Prof. William M. Gray, Colorado State University
Prof. Paul. W. Mielke, Jr., Colorado State University
Prof. Kenneth J. Berry, Colorado State University
Dr. John A. Knaff, Colorado State University
Mr. Rick Taft, Colorado State University


To investigate statistical seasonal forecasting methodologies and to issue real-time seasonal predictions for various tropical meteorology/oceanography phenomena, including Atlantic basin hurricanes, African rainfall and the El Nino/Southern Oscillation phenomena.


The understanding of climate variations on a year to year basis has potentially large positive impacts for communities around the world. To be able to say, with even a modest degree of skill, whether - for example - the coming rainy season will be quite wet or whether a drought may be expected instead, would be quite valuable in making preparations for such events. The meteorology and oceanography in the tropical latitude belts can be seen as fortunate in that seasonal forecast skill does appear to be possible up to a year or more in the future, as opposed to the mid-latitudes where little skill is available. Thus it is logical to take advantage of such a situation and pursue avenues of predictability that can be both useful and of interest to the public community.


Research is first undertaken to reveal which physical phenomena can be utilized as skillful predictors for the seasonal parameter that one wishes to make forecasts of. This preliminary work involves analyzing composite fields of the potential predictors, running Empirical Orthogonal Functions to isolate the various modes of variability, and utilizing partial correlations to remove those predictors which are simply mirroring the associations found in other predictors. Once a pool of potential predictors have been isolated, then a combination of the Least Absolute Deviations (LAD) and the Ordinary Least Squares (OLS) regressions are run to obtain the best fit of predictors to predictands. Estimates of future skill are obtained either by running cross-validated regressions over the predictors/predictands or by using an empirical estimate of the skill shrinkage on future (independent) data. The various prediction schemes should then be run in real-time (and provided to those who are interested) as independent testing is essential to give true estimates of skill.