Browsing by Author "Yihenew, Fekadu"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Constructing predictive model using Data mining techniques in support of Motor insurance policy risk Assessment: the case of Ethiopian Insurance Corporation (EIC)
    (A.A.U, 2015-06) Yihenew, Fekadu; Dereje, Teferi (PhD)
    In recent years data mining has attracted a great deal of attention in information industry due to the wide availability of huge amounts of data and the need to change them into useful information and knowledge for broad applications including market analysis, business management, decision support, risk assessment and fraud detection. This research applies data mining techniques in support of motor insurance risk assessment at the time of underwriting. The research is implemented using the six-step suggested by Cisos et. al.(2000) DM process model. The data collection process has been done in two phases. Records about vehicles are collected from INSIS database where as records about drivers’ are collected manually. The collected dataset is preprocessed using weka DM tools and Microsoft Excel in order to select attributes, derive new attributes, handle missing values and remove outliers. In this study an attempt was made to apply data mining clustering and classification algorithms. K-means clustering algorithm is implemented to come up with the natural group of the claim records as low risk, medium risk, high risk and very high risk. The researcher implemented two classification algorithms, J48 decision tree classification algorithms and multiperceptron ANN. Using j48 decision tree classification algorithms different experimentations are conducted. The first experimentation with default parameter values and 10 fold cross validation test options has registered 94.63% accuracy. ANN experimentations have also been conducted. Accordingly the experimentation with default parameter values with 10-fold cross-validation test option registered 99.58% accuracy. The study also registered an accuracy of 93.74% with percentage split test option by splitting the dataset into 80% to training set and 20% to test set. The result of this study indicates that applying data mining to classify insurance customers to predict the risk level is very promising. The above prediction accuracy also indicates that data mining is a powerful tool to measure the uncertainty of losses in EIC. Hence, the research identified future research direction in order to implement applicable system in risk assessment process.