A statistical scheme to predict the weekly Tropical Cyclone activity in the Southern Hemisphere is developped. Tropical cyclone data are available from two datasets. Ones covers the period 1969-1998 in whole Southern Hemisphere region, the other covers 1969-2002 in the South Pacific area. The quantities predicted are the probabilities that a cyclone forms or occurs over an area during a week. By definition, TC formation happens when a TC central pressure first reaches the value of 995hPa whereas TC occurence happens when a TC central pressure is lower than of 995hPa and its latitude is lower than 30 deg S. The Southern Hemisphere was divided in four different areas: the South Western Indian Ocean, the South Eastern Indian Ocean, the southwestern Pacific Ocean and the Central Pacific Ocean.
As a first step, the effects of potential predictors of this scheme were investigated and index that properly represents these predictors were found. The hypothesis upon which the forecast model was based was that there are certain large-scale influences on TC activity that have some memory and/or predictability. Such possible large-scale influences include interannual variability such as El Nino Southern Oscillation (ENSO), the seasonal cycle, the Quasi-Biennal Oscillation (QBO), the MJO and perhaps others. The influence of the MJO on TC activity was shown using a real-time index of the MJO. The climatological probability that a cyclone occurred or formed was found to be a good index of the position within the cyclonic season. The interannual variability associated with SST, i.e. ENSO and the Indian Ocean variability were also found to affect TC activity. Indices of the QBO, however, were not found to be useful in the model.
The statistical model used is a logistic model which hypothesis were successfully checked. Among the firstly investigated predictors, the choice of those that were included in the model was done by a stepwise selection procedure. Then, the performance of the model in each area was shown, using strictly cross-validated forecast made in the conditions of real-time with climatology recalculated over each set of learning period. The forecast is unbiased with most of the forecasted probability in a 10% interval centered on the observed probability. It's Brier score are between 0.1 and 0.2. Taking into account the MJO significantly improves the forecast of the two first weeks. The interannual variability showed an almost constant potential to predict TC activity up to the seventh week.
Finally, possible improvements to the scheme were considered. Using the last data available gives an almost as good forecast as using lagged variables. Taking into account Equatorial Rossby wave could increase the performance of the model at the first lead especially in the Southwestern Pacific Ocean.