SAS Data Mining & Neural Network as powerful and efficient tools for customer oriented pricing and target marketing in deregulated insurance markets Stefan Lecher, Actuary Personal Lines, Zurich Switzerland Abstract As Zurich Insurance Company makes the transition from a product orientation to a provider of solutions for clearly defined customer groups, the detection of profitable target markets and the specification of solutions which meet customer s financial protection needs became one of the major tasks. Traditionally the pricing of insurance was mostly based on the product and not on the insured, whereas prices in deregulated markets have to be closely related to the each individual customer s exposure to risk. Deregulation and highly competitive insurance markets force insurance providers to increasingly follow this path. Zurich Switzerland uses Neural Networks to detect policyholder s attributes, which have a significant effect on the probability of higher losses or more claims. The results can not only be used to determine risk-adequacy, but also to allow marketing to identify potentially profitable customers. The success of the Neural Network approach has lead to a broader Data Mining project to determine each customers profitability depending on the net present value of future cash flows. The goal of this project is to bundle the marketing efforts and to improve customer oriented pricing by taking into account the probability of future developments. The possible gains of using Data Mining techniques are large, but even larger are the opportunity costs of letting data bases rich of information go untapped, because this data can give us a deep insight into our customers needs. Zurich Switzerland and the deregulation of the Swiss insurance market Deregulation and the ongoing globalisation of the insurance markets in the past years had a significant impact on business development in Switzerland. But those developments not only served to intensify competition and encourage structural change in insurance, by enlarging the entrepreneurial playing field, they also created new market opportunities for tailoring services more closely to customer wishes. This development is even more important in view of the market change in customer demands: Today s customers have access to detailed information on insurance and are looking for ways to satisfy their individual needs. They expect solutions which provide them with value added services. At Zurich, the emphasis is placed on specific problem-solving and innovative provision of customised solutions, not on individual products and therefore new approaches and methods have to be made. To understand local market conditions and customer requirements marketing is done in the Zurich Group through strategic business units (SBUs). Each SBU focuses on specific market
segments or targeted customer groups. The market strategy places special emphasis on establishing long-term relationships with customers, based on a thorough understanding of their needs and the provision of value-added solutions. The selective marketing policies are aimed at specific customer segments and are supported by the latest technology to recognise the new market opportunities and possibilities for increased profit. With deregulation and globalisation competition keeps the market in a steady flux. Services, structures, and operations had to be customised closely to targeted segments to meet customer needs. Zurich Switzerland reacted fast and took strategic and organisational measures. In 996, Zurich undertook several strategic steps in the Swiss market. The non-life and life business in Switzerland was consolidated into one business division under common management, effective April, 996. In September 996, Zurich Switzerland was reorganised around common customer groups and the new organisational structure was substantially in place by April 997. The Swiss market is now divided into customer segments, and the recently created SBUs (for large accounts, small and medium sized companies, financial services and personal lines) are dedicated to these customer segments. Customers can now receive both life and non-life products and financial solutions from a single source. Operating as one of the Zurich Group s 0 SBUs, Zurich Switzerland Personal Lines focus lies on selected market segments of targeted customer groups. Central to the strategy is the strong customer orientation geared towards offering excellent solutions for financial protection. The Care strategy was lunched under the name Relax and is geared towards personal customers and includes the Help Point service, a toll-free hot line serving Zurich customers around the world in all types of situations, twenty-four hours a day. The main part of Personal Lines non-life operations is the automobile portfolio. The deregulation of the automobile third party liability market in Switzerland during 996, previously the industry s most heavily regulated sector, created intense competition, rate cutting and therefore reduced profit margins. Target marketing, customer orientation and monitoring profitability therefore became a unified effort. Business questions in SBU Personal Lines For Zurich Switzerland the strategic direction was set, the opportunities in the deregulated markets recognised and work could start in the SBUs to target their customers. SBU Personal Lines has relationships with around 800'000 customers in Switzerland. Gross premiums in 996 amounted for US$. bn, which is a market share of roughly 0 %. Before the strategic change Zurich segmented the market by products, so a policy with a defined coverage was the main unit of analysis. E.g. in Actuarial and Marketing the customer was defined by having a auto insurance policy. If the same customer had a life insurance policy he was considered a different customer. The focus now lies on the integral customer which generated a lot of new business questions were addressed Actuarial by Marketing Departments. Differentiated solutions for different customer groups that need to be developed and priced, the needs of our customers for integral protection have to be addressed, the pricing has to be changed to give credit to low risk groups, and customer groups have to be targeted in different ways according to their needs. To guarantee Zurich s commitment to building up strong and trustworthy relationships over
time and to firmly stand for financial stability and value creation, the long term profitability of the developed solutions has also to be secured. To sum up the main question: What are our customers needs, which are the solutions that cover those customer wishes and which of the solutions are competitive and profitable? Therefore focus shifted from the statistics of policies to the analysis of customers. The rest of the paper will concentrate on two examples, which will show how SBU Personal Lines of Zurich Switzerland addresses the main question. First, the auto insurance tariff was completely revised with the deregulation of the market in 996 to meet customer s needs. Second, with the organisational restructuring and focus on private customers, target markets have to be defined, different products have to be reshaped to provide solutions for differentiated customer groups and a risk-adequate pricing has to be steadily improved to guarantee long term relationships and financial stability. Customer oriented auto insurance pricing The main purpose of the new motor tariff was a pricing of the policies according to every individual customer s exposure to risk. Allowing tariffs in deregulated markets to vary between risks with different probabilities to produce a loss, customers would no longer accept the same rate for all drivers. To meet these needs and fulfil the promise to provide a customer oriented solution, a modular product was developed. The client can now choose the coverage, deductibles and limits he wants. Additionally, homogeneous groups had to be determined and for each of these groups the expected losses were calculated. Here homogeneous classes mean that in every risk class the probability of succeeding a claim and the expected average claim size are as equal as possible. The main question now was, which combination of attributes constitute a homogeneous risk class. The problem was solved by two approaches: first by classical statistics and second by using neural networks, a machine-learning algorithm and important data mining technique. In the statistical approach theoretical assumptions and business knowledge constituted the starting point. Those attributes believed to be important to determine the homogeneous risk classes were used to construct the groups. Data of the past five years was grouped in these classes and the average loss ratio (losses divided by premium) was calculated. The deviation of the loss ratio of every risk class to the average loss ratio of the whole portfolio constituted the rating factor in the tariff. Consider the age of the driver in groups and the driving experience in 8 groups, this creates 0 classes. For each risk class the deviation from the mean was calculated (figure ). Because the variation between different classes was large and some classes consisted only of a few risks, a credibility model was applied and large losses were capped to smooth the factors (figure ). Finally the factors were adjusted by hand to guarantee reasonably downward adjustments of the premium as a customer gets older and becomes a more experienced driver (figure ). In the second approach SAS Neural Network was used to model the risk classes. Different network configurations were tested. To keep the model simple just the most significant
attributes were included. Finally the neural network produced similar results as the statistical approach (figure ). Figures to : Loss ratios of risk classes (grouped by age and driving experience) compared to the average loss ratio Age class 7 6 Experience class Age class 7 6 Experience class Figure Figure Age class 7 6 Experience class Age class 7 6 Experience class Figure Figure Following two approaches to finally get almost the same results may seem an unnecessary luxury. But when you change the pricing structure of a large portfolio you can never be sure that you considered all relevant effects. By obtaining similar results by two independent methodologies added credibility to the process. Comparing the approaches two points can be mentioned. First the neural network does require less assumptions about how the attributes relate to the loss ratio. Second, much less ad-hoc programming is necessary. With SAS Neural Network the modelling can be done by graphical user interface. Whereas with the credibility modelling, with the capping the losses
and with the final adjustments, a large amount of SAS code had to be written. So time savings is an obvious and also an important benefit of the neural network approach. On the other hand, the basics of classical statistics are widely understood by non-experts and are easier to explain to other people involved in the product development, whereas explaining the insight of a neural network is almost impossible. Nevertheless applying neural networks can give you additional insights into relationships, which can assist you in understanding the business problem. Benefits of risk-adequate pricing The new tariff structure not only brings individually rated automobile insurance policies. With the implementation of a sensitive bonus-malus system additional credit is given to the individual loss experience of each driver. The customer service was also highly improved by the sales support process, which allows the agents on-site production of sales documents and direct data transfer with laptops. This process not only serves the customer through on the spot policy issuance but also guarantees a fast feedback of information into the analysing and monitoring process. Two full years of experience with the new tariff structure now show astonishing results. First, the loss ratios of almost all the differentiated homogeneous risk classes correspond to the calculated premiums, so that every risk class shares the appropriate portion of the whole losses and customers are therefore priced fairly. In the long run this will be honoured by our customers as no one has to subsidise groups with higher average losses. Second, the new tariff structure changed the weights of groups in the portfolio. With deregulation people started to shop around for lower premiums. This effect can be seen in the Zurich auto portfolio as before subsidised high risks declined, whereas low risks took advantage of the risk-adequate premiums. E.g. Zurich grew faster in the segment of elderly drivers (groups to ) and grew slower in the segments of high risk drivers (group ) (see table ). Table : Distribution of new customers in risk classes 00% 80% 60% 0% 0% 0% % % 0% % % % % 9% Old Tariff New Tariff & Both effects improve the profitability of the whole portfolio, which allows further riskadequate rate cuttings and guarantees financial stability at the same time. The success of the two way approach by classical statistics and the new data mining technique paved the way of future pricing. To further improve the pricing of homogeneous risk classes,
in the future not only neural networks but also decision trees besides multivariate statistics will be used to develop, monitor and implement even more complex models, to offer customer oriented individually rated policies, which lies in line with customer s wishes and the profitability goal of Zurich. Target marketing by detecting profitable market segments The experience made with data mining techniques in actuarial pricing of auto insurance policies led to an ongoing project to define profitable target markets for which customer oriented solutions will be developed. First, the customer s needs have to be researched. We concentrate on questions like which customers demand which coverage and which combinations of products. The answer to this question allows on the one hand to target every customer with the desired product and improve cross-selling, and on the other hand to develop tailored and differentiated solutions with combinations of products for financial protection for defined customer groups with similar needs. Thereby the company s resources can be channelled to those customers which desire integral long term financial protection. This can mean that younger customers may purchase only small basic coverage but develop other needs, e.g. as they get older, their income changes, they get married, etc. Customer oriented solutions in this sense mean steadily adapting of the financial protection according to the changing needs. Second, the profitability of a segment of customers with similar needs has to be determined. Traditionally in pricing, accounting and management premiums and losses of a one-yearperiod are compared. To calculate profitability one has to include costs in a explicit way, because costs for administration, overhead, commissions, etc. are treated separately from the payments for losses. The largest part of the costs occur at the beginning of a relationship with a new customer. Therefore, in the first year a customer never would be profitable. That s why the expected future development of the customer relationship has to be taken into consideration and costs have to be distributed in some way over the whole duration of the relationship. A basic model of the customer profitability has to be developed by taking the net present values of all future cash flows. For one product, e.g. auto liability insurance, we have to estimate future premium payments, future loss payments, and the costs which arise at different points of time. In addition the probability of purchasing additional coverage or cancelling the policy have to be incorporated in the model. The model gets even more complex if we want to include cross-selling aspects. Then the probabilities of future purchases of additional products have to be estimated and profitability models of those additional products must be included. Such an extensive model depends on many assumptions and estimated parameters, so that results have to be interpreted with caution. Nevertheless the development of such a model will give us deep insight into our customers needs and customer relationship. But for sure target marketing, customer oriented pricing and establishing long term relationships, which are all cornerstones of profitable business, can be done in a better way if one keeps in mind the limits of the used model.
Model building using SAS software and first experience with SAS Enterprise Miner To build such an ambitious model different steps are taken. As we still are in the ongoing process of testing different model, final results are not yet available, but some initial observations were made. Future cash flows of premium, losses and costs have to be estimated. The premium payments are the easiest part as they are fixed and depend on attributes of the insured and/or the insured object. Forecasting individual losses is difficult, because by nature they are stochastic. An expected average has to be estimated depending partially on the specific individual customer and the whole portfolio. Finally assumptions about fixed and variable costs per customer and per policy are included. Historical data is used to estimate parameters and make assumptions. To prepare the data, to clean the data, to examine the data, to do descriptive statistics and finally estimate parameters SAS/Base programming language, SAS Insight (both accessible directly through a SAS Enterprise Miner node) and the beta-release of the SAS Enterprise Miner are currently used. The final goal is to get a file with one record per customer with information about the customer including the estimated profitability, information about the insured objects, information about selected coverage, and information about the expected losses and costs. On this file data mining techniques will be applied to find relationships between customer attributes, insured objects, selected coverage, and profitability. To do this analysis SAS Enterprise Miner allows us to use classical statistics (clustering, associations, regression) and data mining techniques (tree based models, neural networks) with the same graphical user interface. Additionally, we can compare and assess the results from different techniques, e.g. a tree based model allows to determine attributes which make it likely that customers purchase additional coverage in motor insurance. By applying the data mining techniques we hope to get new insight into the relationships between customer attributes and customer needs, which can not be discovered by examining our data by visual exploration and by classical statistics. Conclusions Changing markets and customers demands forces insurance companies to be flexible and respond to the new challenge. To know your customer becomes a central point and now marketing and actuarial are working closely together to detect customers needs and develop tailored solutions which are competitive and profitable. First results in risk-adequate pricing and modelling customer relationships are promising. We still are making the first steps in a new land. New techniques with new algorithms demand new knowledge, which combines business and mathematical understanding. Software tools, as SAS Enterprise Miner, can facilitate the access to these techniques. The graphical user interface of SAS Enterprise Miner allows an easy handling and fast start, but those are not a substitute for insufficient knowledge and understanding of business questions, the modelling, the data and underlying technique.
Building complex models demand a lot of simplifying assumptions, many parameters have to be estimated and the model specification is always a problematic task. SAS Enterprise Miner does not solve all these problems but helps you structure and document the research process, and makes it easier to understand and apply the new techniques. If you are already familiar with other SAS products, the full compatibility is given, so that you can build on a already existing stock of knowledge. As we are still in a testing process the efficiency evaluation is not yet done. The first experience shows, that significant time savings can be achieved and that new questions won t be answered without applying data mining. So the main question is if we can allow to leave rich data bases with information about our customers unused. Probably the opportunity costs of not starting now can become very large.