Predictive Modeling Using Data Mining Techniques in Support of Insurance Risk Assessment
No Thumbnail Available
Date
2002-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
One of the important tasks that we have to face in a real world application is the task of
classifying par1icular situation or events as belonging to a certain class. Risk assessment in
insurance policies is one example that can be viewed as classification problem. In order to
solve these kinds of problems we must build an accurate classifier system or model. Data
mining techniques are powerful tools in addressing such problems.
This research describes the development of predictive model, which determines the risk
exposure of motor insurance policies. Decision tree and neural network were used in
developing the model. Since rejections of policy renewal are rare at Nyala Insurance SC.
(NlSCO), where the research was conducted as a case study, policies were classified into one
of the three possible groups (Low, Medium, or High risk) on the basis of annual assessment
made by NISCO. Six variables were extracted from the 25 variab les used in this study. 940
facts (90% of the working dataset) were used to build both decision tree and neural network
models. The remaining 116 (10 %) of the dataset were used to validate the performance of the
models. The decision tree model, selected based on the meaningfulness of the rules extracted
from it, correctly classified 95.69% of the validation set, and the classification accuracy for
low, medium and high risk policies are 98. 15%, 94.12%, and 92.86% respectively. The neural
network model correctly classified 92.24 % of the validation set, high-risk groups are
correctly classified, and low and medium-risk groups are classified with accuracy of 98 .1 5%
and 76.47% respectively. Some possible explanations for the relatively low performance of
the neural network with medium policies are given. In addition, an interesting pattern was
found between the two models that some policies classifieds by decision tree were correctly
classified by neural networks, and vice versa. This is a good indication that the hybrid of the
two models may result in better performance.
Description
Keywords
Information Science