Page 363 - Contributed Paper Session (CPS) - Volume 6
P. 363

CPS1966 Jessa L. S. C. et al.
            has  a  higher  weight,  which  may  also  indicate  a  decline  in  health,  Other
            Traditional is more often bought. In terms of income, Unit-linked can also be
            preferred by low earning individuals.
                To assess the accuracy of the models in predicting, the final models were
            used to predict the test data.

                   Table 3: Summary of Misclassification Rates for All Models (Test Data)
                                            Total Misclassification     High-risk
                                                                     Misclassification
                 Classification Trees            24.37%                  12.83%
                 Multinomial Logistic
                 Regression (CART Final          17.52%                  12.08%
                 Predictors)
                 Random Forest                   18.81%                  15.55%
                 (Full Model)
                Among  the  models,  only  the  final  models  from  Multinomial  Logistic
            Regression  and  Random  Forest  passed  the  misclassification  benchmark  of
            20%  assuming  total  misclassification.  However,  if  the  high-risk
            misclassification  was  considered,  which  is  the  portion  of  Endowment  and
            Other Traditional observations classified as Unit-Linked, Classification Trees
            can  also  be  accepted  at  the  threshold.  The  high-risk  misclassification
            represents the conservative clients which may be offered with an aggressive
            investment insurance product without caution.

            4.  Discussion and Conclusion
                All models are acceptable within 20% misclassification threshold if only the
            high-risk misclassification is considered. The models are also shown to be poor
            in predicting minority classes  but good in classifying the dominating class
            which is Unit-linked. Given these results, the models cannot be used to profile
            the  markets  for  Endowment  and  Other  Traditional  but  were  used  to
            understand  the  market  for  Unit-linked  better.  The  relationships  that  were
            generated from each of the models can be consolidated to come up with a
            targeting  strategy  to  promote  Unit-linked  plans.  Tapping  the  Endowment
            market as well to purchase Unit-linked is reasonable since the two products
            address the same need which is savings and that selling Unit-linked is more
            profitable for the company given the instability of the interest rate over the
            long-run.
                Caution must be observed however in offering Unit-linked plans using the
            insights from the model since much observations from the minority classes
            are misclassified. Offering an aggressive financial tool to conservative clients
            might cause severe misselling. This risk can be controlled by full disclosure of
            the features of a Unit-linked product especially its non-guarantees.
                There were a lot of insights that were derived from the models that can be
            used for the targeting strategy. For instance, Unit-linked plans can be offered


                                                               352 | I S I   W S C   2 0 1 9
   358   359   360   361   362   363   364   365   366   367   368