A Data Mining Model for Risk Assessment and Customer Segmentation in the Insurance Industry

Customer segmentation on the basis of predictable risks can help insurance firms maximize their earnings and minimize their losses. Car insurance is one of the most lucrative and profitable branches in the insurance industry. Utilizing the concept of self-organizing map, the authors propose a two-phase model called ‘Auto Insurance Customers Segmentation Intelligent Tool’ to segment customers in insurance companies on basis of risk. In the first phase, the authors extract 18 risk factors in four categories consisting of demographic specifications, auto specifications, policy specifications, and the driver’s record extracted from the literature review. In the second phase, they finalize the selection process by drawing on expert opinion polls. The authors utilize self-organizing maps since they are able to display the output in the form of illustrative and comprehensible graphical maps capable of representing linear and non-linear relationships among variables, insensitive to the learning input, and slightly sensitive to the noise in the learning input. Finally, K-means are employed to compare the results with those obtained through self-organizing maps. DOI: 10.4018/jsds.2013010104 International Journal of Strategic Decision Sciences, 4(1), 52-78, January-March 2013 53 Copyright © 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. derstand customers’ desires, allocate resources, and support them against competitors (Dickson, 1982). Customers vary in many essential ways. Generally, customers can be distinguished from each other from two different views, that is, from their value for the organization, and the difference in their needs (Hosseinzadeh, 2008). Cycle of a customer relationship management (CRM) consists of four dimensions. CRM begins with customer identification. This phase involves targeting the population that is most likely to become customers or most profitable to the company. Elements for customer identification include target customer analysis and customer segmentation. Target customer analysis involves seeking the profitable segments of customers through analysis of customers’ underlying characteristics, whereas customer segmentation involves the subdivision of an entire customer base into smaller customer groups or segments, consisting of customers who are relatively similar within each specific segment (Ngai et al., 2009). Risk segmentation is, in fact, the segmentation of customers with similar risk characteristics, probably causing similar damage. Risk segmentation of policyholders on the basis of observable features can help insurance companies to reduce loss, raise the rate of insurance coverage, and prevent them from making an inappropriate choice in the insurance market. The majority of developed and developing countries take advantage of the risk segmentation in determining the insurance premium in their insurance industry (Majed, 2008). The rate of the insurance premium in many insurance companies is computed with regard to different demographic variables, car specifications, and the record of damage caused by the car owner. Mike Kreidler, a member of the board of directors of one of the insurance companies in Washington State, pointed out that the rate of insurance premium relies on factors such as the policyholder’s age, gender, and marital status, vehicle type, the location of the car owner’s residence, the claim history, and the driving pattern (Kreidler, 2008). While in Iran the rate of comprehensive auto insurance premium is set by Central Insurance of Iran. In practice, low-risk customers pay for the damage and loss caused by high-risk customers, so there is no difference between these two groups of customers. In addition to the inefficiency of the contracts of insurance or policies, lack of such measures in determining the risk in automobile insurance leads to computing unfair rates because in these cases instead of the person, the car is insured. That’s why most insurance companies experience great loss as far as automobile insurance is concerned while most developed countries have attempted to increase the productivity and profitability of their insurance industry using the risk segmentation system (Hosseinzadeh, 2008). Self-organizing map (SOM) is a type of artificial neural network that is trained using unsupervised learning to produce a representation of the output in the form of visual graphical maps which are comprehensible to managers of organizations. The insensitivity of the selforganizing map to the learning input, and its slight sensitivity to the noise in the learning input, its capability to display linear and nonlinear relationships among variables, and its great power to segment data are among the advantages of these maps. In this research, this map in data mining is used to segment customers of automobile comprehensive insurance using the recognized factors in risk. Since this map is able to illustrate the output in the form of graphical maps, it is easier for managers and experts to understand and interpret the results. In the present study, the important and effective factors in the risk of policyholders will be identified and picked out in two separate phases. In the first phase, the study and review of the articles published in the prestigious journals including ‘Science direct’, ‘IEEE’, ‘Emerald’, ‘ProQuest’ and scientific reports within a span of 9 years from 2000-2009 was carried out to select risk factors. The extracted factors were generally grouped into four main categories, that is, demographic specifications, automobile specifications, insurance policy terms, and the driver’s record. In the second phase, a questionnaire concerning the identified factors was designed and filled out by experts from among those working in financial compensation department and automobile comprehensive 54 International Journal of Strategic Decision Sciences, 4(1), 52-78, January-March 2013 Copyright © 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. insurance administrators in ten insurance companies. Finally, after analyzing the responses to questionnaires, using Scree test, the final risk factors were chosen. To obtain the required data, the database of automobile comprehensive insurance in Mellat Insurance Company (MIC), as the biggest and the first electronic insurance company in Iran, was utilized. MIC is one of the active and successful private insurance companies which began working in 2003 in Iran. It comprises around 400 agencies countrywide and enjoys the highest information technology. After segmenting the insurance policyholders using the proposed SOM, as a rival method, K-means, a segmentation technique, was employed to assess the reliability of the proposed method. This research consists of five sections. The introduction is followed by Section two, the review of literature, which in itself comprises three sub-sections presenting customer segmentation, SOM and K-means algorithm concepts. Methodology is discussed in Section three that separately describes the phases of the proposed model. Conclusion is discussed in Section four and finally suggestions for further research are offered. 2. REVIEW OF LITERATURE In the CRM cycle, the first and foremost step is to recognize customers. Since the customer is viewed as the major factor in businesses’ income and profitability, how to interact with them has a deep impact on the amount of profit a business can make (Choobdar, 2008). At the same time, customers differ in many fundamental ways. Generally speaking, customers can be distinguished from two different views, that is, from their value for the organization, and the difference in their needs. Such distinctions make it possible for organizations to improve their interactions with customers through segmentation and exploit their energy and resources in an appropriate and wise way (Hosseinzadeh, 2008). Segmentation has been identified as one of the most significant concepts in marketing. Skillful segmentation helps firms to pick out profitable customers, understand customers’ desires, allocate resources, and support them against competitors (Dickson, 1982). If an organization is not able to obtain and understand the minimum characteristics of its customers, it will miss out the opportunity to implement its strategic plans (Beane & Ennis, 1987). 2.1. Data Mining and Customer Segmentation Due to the sharp rise of the information technology (IT), the amount of data stored in databases is dramatically on the rise. Analyzing the stored data and converting it to information and knowledge which is applicable in organizations requires powerful instruments. Data analysis is an influential tool that has been widely employed to extract information and search for relationships and patterns in the huge amount of data. According to the studies done so far, various data mining techniques have been used to segment customers, such as decision tree technique (Kim et al., 2006), K-means algorithm (Dennis et al., 2001), pattern-based clustering (Yang & Padmanabhan, 2005) and self-organizing maps (Kim et al., 2006). In the CRM cycle, segmenting and analyzing customers is the essential step in recognizing customers. Data mining techniques employed so far in customer segmentation can be divided into two main categories which include customer classification and clustering (Ngai et al., 2009). According the study done by Ngai et al. in 2009, it was found out that from among the 34 techniques utilized to mine data, neural networks are the most commonly used. This article states that from among 87 articles published on customer management in 24 prestigious journals during 2000-2006, in 30 articles, that is, 34.5% of all articles on customer management, neural networks technique has been employed. Of the 30 articles which deployed neural network techniques, 16 (53.3%) adopt SOM subtypes, which entail mapping structured, high-dimensional data onto a much lower-dimensional array of neurons in an orderly fashion through the training process (Ngai et al., 2009). Table 1 indicates the studies done on the customer recognition. International Journal of Strategic Decision Sciences, 4(1), 52-78, January-March 2013 55 Copyright © 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited. 2.2. Techniques Adopted in This Research 2.2.1. Review of the SOM SOMs are subtypes of neural networks, which are trained using unsupervised learning, and are able to analyze complex spaces. SOM algorithm was first introduced by Kohonen in 1981. These maps simulate the human beings’ retina nerves and were first practically used in 1984 for voice recognition (Kohonen, 1984). Although SOM is frequently used in data mining, complex spaces display (Vesanto, 1997), clustering the spaces with high dimensions and particularly in image processing, project management, financial analysis and industrial detections and medical diagnoses, comparatively it is less used in managerial and business administration related fields (Oja et al., 2002). An extensive list of engineering usages of SOM is presented by Kohonen et al (Kohonen, Oja, Simula, Visa & Kangas, 1996). The basis of SOM is to map spaces with high dimensions into twoor three-dimension space in a way that minimum information is lost and the hidden information in relations among the data can be discovered and showed. This method shows the correlation between data, information and their mutual effects on each other (Hanafizadeh & Mirzazadeh, 2010). The SOM network typically has two layers of nodes, the input layer and the map layer. Each map includes a set of neurons that are put together in a two-dimensional grid which is fully connected to the input layer. Each neuron of the map layer is corresponding to an information vector with the dimension numbers equal to the dimension number of the information space under analysis. In other words, each neuron is the representative of one part of the information space. Figure 1 presents topology of SOM. 2.2.1.1. Training SOM The SOM basic algorithm relies on a competitive unsupervised learning algorithm known as “winning takes all”. The network undergoes a self-organization process through a number of training cycles, starting with randomly chosen weights for the neurons in the map layer. During each training cycle, every input vector is considered in turn and the winning neuron is determined. SOM training algorithm involves the four following steps) Kohonen, 2001(: Table 1. Different techniques of data mining employed in identifying customers (Ngai et al., 2009) CRM dimensions CRM elements Data mining functions Data mining techniques References