本研究利用資料採礦技術中的邏輯特迴歸與類神經網路來預測壽險保單早期失效之機率。採用Boritz and Kennedy(1995)的誤判成本來衡量預測模型的優劣性,研究結果發現,運用倒傳遞類神經網路可以有效地解決邏輯特迴歸以驗前機率做為臨界值而產生的型I錯誤過大的問題,並降低誤判的損失成本。其次,利用邏輯特迴歸所篩選的顯著變數,可以解決倒傳遞類神經網路無法說明變數間的因果關係之問題。除此之外,倒傳遞類神經網路之誤判成本並不會隨著成本率的增加而有太大的變化。最後,經由實証得知業務人員是否跳槽或離職、服務品質、繳費方式與保費負擔能力等因素對保單是否早期失效有很顯著的影響。
This paper applies data mining technology to predict the probability of early lapse for life insurance policy. Based on the misclassification cost of Boritz and Kennedy (1995), we compared the performance between the neural network and logistic regression models. The results of the study indicate that, firstly, back-propagation neural network can solve the problem of logistic regression that use prior probability for cut-off value that result a higher type I error. Secondly, by using the significant explanatory variables obtained from the result of logistic regression, it is helpful to explain the relationship between output and input variables of back-propagation neural network model. Moreover, misclassification cost of the back-propagation neural network is more stable than others in varied cost ratios. Finally, the most important factors that affect the probability of early lapse are the problem of salesmen's quitting job, the service quality of business affairs personnel, premium payment mode and ability to pay premium.