Prospects for Improving the Operational Seasonal Prediction of Tropical Cyclone Activity in the Southern Hemisphere - Monitoring and Prediction of Tropical Cyclones in the Indian Ocean and Climate Change - page 129

eastern South Pacific Ocean the strongest relationship with the number of TCs
was found for September indices, although the association was weak (0.297
for the 5VAR, 0.318 for the NIÑO3.4 and 0.273 for the SOI). With these weak
correlations, we conclude that it is not sensible to further build linear regression
models.

3.3 Statistical-dynamical Model-based Approach
for TC Seasonal Prediction

The Australian Bureau of Meteorology has developed a dynamical climate
prediction model POAMA (Predictive Ocean-Atmosphere Model for Australia)
(Wang et al., 2008). It has been demonstrated that POAMA has substantial
skill in predicting SSTs and rainfall across the Asia-Pacific region (Hendon et
al., 2009). The skill results indicate the potential for developing TC seasonal
prediction using statistical-dynamical model-based approach. Developing the
improved statistical regression models, we found that the 5VAR index
demonstrated the best correlation with the TC number and the correlation was
constantly high (close to 0.7 in the Australia region) for six months from August,
A( t ), to January of the next year, J( t +1) (Section 3.1). It leads us to a proposition
that, using outputs from the dynamical model POAMA obtained prior to the
TC season, it is possible to compute predicted values for the ENSO indices in
advance (e.g. POAMA outputs generated in August and September can be
used to compute October-November-December 5VAR values) and then use a
statistical model to predict TC number. This approach may be particularly
beneficial for early warning of expected active TC season (issued 1-2 months
ahead of the statistical model-based forecast). In this study, we simulated such
approach by the development of a regression model using October-November-
December values of the 5VAR index as a predictor. Similar to the results in
Section 3.2, the model with additional time trend has better performance than
having 5VAR index only as the single explanatory variable.
There were four observations flagged as potential influential points in our
proposed model, and they were observations 1, 20, 29 and 39. The models
were re-built iteratively each time by omitting one flagged observation, and
the model that excludes the first observation (1969-70 season) was retained
due to its highest adjusted R 2 value (0.4932). The results are given in Table 5.
One can see that both the 5VAR index and time trend are significant at 5%
level of significance in the model.

Table 5: Multiple linear regression with both time trend and the
October-November-December 5VAR

Adj R 2

T 5VAR Residual SE

Coefficient -0.086 -1.935 2.410 0.493

P -value 0.014 6e-06