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 127

Table 3: Multiple linear regression models (as per Table 2)
with influential points removed

Model 1.1 Model 2.1 Model 3.1

Observation 1 1 1

removed

Coefficients p-value Coefficients p-value Coefficients p-value

T -0.079 0.026 -0.083 0.025 -0.094 0.007

5VAR -2.055 8-06 - - - -

NINO3.4 - - -2.350 4e-05 - -

SOI - - - - +0.218 4e-06

Residual s.e 2.428 2.528 2.391

Adj R 2

0.486 0.443 0.501

region, a reduction in TC activity is usually observed in El Niño years, while
in La Niña years TC activity is typically higher compared to El Niño years
(Nicholls et al., 1998; Kuleshov et al., 2008). In 1970s, four La Niña events
and four El Niño events were identified, while during the next three decades
only five La Niña events were observed, with 12 El Niño events recorded
(Kuleshov et al., 2009). Such distribution in frequency of the ENSO cold and
warm phases is one of the plausible reasons for the observed downward trend
in TCs in the Australian regions. This trend could also reflect a slow periodicity
in TC variability due to variation in major climatic drivers with a period rather
longer than the study period (see, for example, impact of the Pacific Decadal
Oscillation on TC variability over the western North Pacific described in Liu
and Chan (2008)). Thus, incorporation of a temporal trend in the statistical
model requires its regular revision, perhaps annually, accounting for TC activity
in the last season and adjusting the model accordingly.
By comparing with the simple linear regression counterparts in Kuleshov
et al. (2009), all the developed models have improved performance. For
example, the model using SOI as a predictor was not the best model in our
earlier study with the adjusted R 2 of about 40%, while in the current analyses
the model demonstrates an improvement in modelling the annual number of
TCs with the R 2 reaching 50%.
Cross-validation was employed to assess the models' performance, each
time leaving one observation out and validating the analysis on that single
observation. The RMSE (root mean squared error) was calculated as the
measurement of fit. RMSEs were 2.72, 2.86 and 2.70 for the Models 1.1
(5VAR), 2.1 (NIÑO3.4) and 3.1 (SOI), respectively. The cross-validation results
are in agreement with those using adjusted R 2 as the model assessment criteria,
that is, the models which used the pre-season July-August-September SOI and
September 5VAR indices and the time trend as the predictors demonstrated the
best performance.