Risk Prediction of Automobiles - Research Paper Example

RISK PREDICTION OF AUTOMOBILES A major part of financial services is risk prediction and analysis and it even encompasses the provision of services for automobiles. There are a number of such provided by almost every financial service that included risk analysis for purposes of insurance, financial loans, mortgages etc. however, as is common case with any industrial sector, there are a number of ways in which the risk is perceived and calculated by every company. As such, this paper is aimed at elucidating the various ways in which financial entities calculate and predict the risk associated with automobiles. In this case, the research will take into account the two different estimate values namely the insurance risk rating and the normalized losses. While the former determines the degree to which an automobile is risky than its price indicates, the latter provides the estimate of the relative average loss of payment per insured vehicle year. One of the most prominent risk analysis methods that are in use today in various insurance companies is the insurance risk factor profiling technique. This is usually achieved by analyzing data that has been collected over a period of time for insured entities. The information collected is segregated according to a number of variables. In the case of automobiles, the variables usually consist of both numerical and categorical data entries and most often; such data may also have noise characteristics. However, the incidence of noise in the analysis and values is minimized by using software profiling tools that help in finding out specific patterns within variables, correlations among different sets of variables and the relationships between a set of variables as the need be (William Mendenhall, 2001). Most often, these tools utilize the help of artificial intelligence logic such as neural networks and incremental learning that in addition to determining useful results, helps use existing data to determine trends for the future. However, the varied efficiency in reducing noise for different sets of data is one deficiency of these risk profiling tools. The use of software tools facilitates the extraction and processing of large volumes of data and requires minimum cost, time and effort (Douglas L. Reilly, 1996). An example of such profiling is as shown below: Fig: risk factor profiling technique (Found at: www.roselladb.com) However, the above method demonstrates the fact that there would be a large difference in the case where the insurance claim would be dependent on a number of factors/reasons. As such, the construction of all possible situations would result in the fact that the results of analysis would provide different results depending on the set of conditions chosen. For example, the consideration of mobile phone (with or without other corresponding reasons) would yield different results depending on whether it were considered on the basis of the first or the second branches starting from the central policies node. Another popular approach that is used by many financial and insurance consultants is the statistical technique of regression analysis. Regression analysis is simply the examination of a dependent variable by using a set of independent variables. The relationship between the dependent variables and the various independent variables in defined by a mathematical regression equation and this is used as the foundation for any regression based analysis. For example, in the current case, a finance company could construct a equation relating depreciation rate to the insurance risk rating and the normalized losses in a suitable mathematical format (consisting of various constants known as regression parameters) and use it to determine the amount of loan that can be granted to an applicant. Though the method is used extensively in these domains given its inherent ability to connect the available data with the values that need to be estimated it is generally perceived that it is sometimes limited in terms of extensive applicability and has also proven to be inflexible at times. Moreover, regression cannot be applied for the estimation of categorical variables and finds maximum use in analysis of numerical variables only. As such, in cases where it is applied for categorical data, the need to perform some transformation is an external burden (William Mendenhall, 2001). For example, consider the use of two variables under the estimation of risk using the risk analysis method by a company. One variable could be the depreciation of value per annum (which is a numerical value) and the other could be the quality level associated with the car (the quality is a categorical value and can consist of values such as 'very good', 'good', 'average', not good', 'bad', 'very bad'). A look at the variables could suggest that one could use the first variable without any hassles for the regression analysis, while the consideration of the values for the second variable 'quality' is only possible is the categorical values were converted into suitable numerical equivalents, for example, very good as 1, good as 2, average as 3 etc. thus, such conversion is an additional burden on the entity handling the analysis. However, once upon conversion, the results are quite qualitative in nature. Typically, in financial and insurance firms, profit generation is usually dependent on the increase in the number of service contracts and/or by a reduction of loss from active contracts. In order to provide for maximization of profits, many more sophisticated models capable of providing accurate results are used. One of such methods is the use of the Cramer decision tree that utilizes statistical estimation and segmentation. The risk prediction models that use the Cramer decision tree utilize the approach of predictive segmentation. Under this approach, the decision is used to segregate the various customers into separate hierarchical segments on the basis of certain attributes concerning the automobile owned by them. Under each such splitting, it is always ensured that any such division is done with the aim of enhancing the policy profiles depending on whether they are safe or too risky. Thus, in the present context, the insurance risk rating could be suitable criteria that could be used for such tree generation. However, this method is not known to work conclusively on certain occasions due to the reason that the number of claims is usually very less. Under such scenarios, classifying all existing policies on the basis of previous claims as being either safe or risky would mean segmenting using highly skewed estimates, which would make it difficult to construct the tree at each successive level as ever new level would increase the level of skew. However, once the tree has been generated, it is quite easy to determine the level of risk associated with the policies by simply examining their probabilities. High probability implies that the group at the level is risky and vice versa (McLeod, 1993). Fig: Cramer's decision tree (Found at: www.roselladb.com) The purpose of demonstrating the above figure is solely for the purpose of demonstrating the decision tree that is obtained by the splitting of available data on the basis of specified criteria. This is evident in the above tree that is supposed to categorize the input data on the basis of jobpost, salary or race. However, if the categorization of criteria were to be considered in the case of risk analysis for automobiles, it will be known that the phenomenon of splitting will be rather skewed in the case of insurance claims as the number of claims would be very less in comparison to the total number of policies. Thus, splitting the decision tree across different levels on the basis of previous percentage of claims versus total number of policies would be rather unbalanced and result in processing at a particular branch of the tree. Thus, this method is an extra burden on the software resources that perform the splitting and in extreme cases, the splitting of the tree at the later levels becomes difficult to process and use for calculation of the input data. Thus, though this method is quite better than the previous methods, it still has some inherent disadvantages. The splitting is however less bothersome in the case of risk prediction in other areas such as loans, mortgage, finance etc. Another popular risk prediction technique is the rule based modeling technique that provides an edge to the analysis over tree-based methods that have certain flaws in their risk prediction efficiencies owing to the skewed nature of data. The rule based modeling technique is implemented by the combination of mathematical rules with models such as decision trees and neural networks. Thus, it can be seen that two different risk estimation methods discussed above can be combined to generate an efficient and robust methods for the same by using rule based modeling. Rule based modeling comes with a number of advantages and is generally regarded as better than all other approaches as it takes the best out of available approaches and uses them appropriately to provide the best results possible (Miley Wesson Merkhofer, 1987). As has been described before, the skewed data can affect the performance of neural networks and decision trees. However, rule based modeling enables the tools to concentrate only on segments that are perceived as high risk zones and allows further analysis to focus on them, thereby providing faster and efficient segmentation and decision making. A rule based model can be considered as an intelligent decision-capable system that can provide answers to complex questions and situations. The decisions are made through a set of rules wherein the system implementing the model is regarded as an expert system. Apart from the rules, a rule based system also works towards updating its intelligence and decision making capability by being in the process of constantly updating itself with facts and events. For example, consider a person who gives information on weather. His/her decision would depend on the following criteria: In case of a rainy day, advice people to use umbrellas when out. If it's sunny, advise people to wear cotton clothes. If it's cold, advise people to remain indoors. The above listings would serve as the rules for the rule based model, the decisions of which would depend on the following facts: 1. It is cold today. 2. It is sunny today. 3. It is rainy today. The facts to the rule based system would be fed through a suitable interface, which would then be able to make decision by comparing the facts against the rules and deciding on the appropriate output. Taking these aspects into consideration, a rule based system possesses the following characteristics: 1. A rule based system that works on rules, facts and a processing logic. 2. Applies rules on facts. Rules denote knowledge while facts denote data. 3. The rules defined contain a set of conditions and actions pertaining to those conditions. 4. A rule may trigger the execution of other rules (Chaur wu, 2004). Rule based modeling also allows the use of third party models that have been developed by other companies and research establishments thereby providing for quick incorporation of newest methods and techniques. Also, insurance companies are occasionally troubled by exceptional cases that usually require them to be considered separately. Rule based models provide for the consideration of these external cases as well as the need arises. Apart from this, the single most utility of rule based models allow for the examination of multiple sets of data thereby helping in arriving at the best possible results that take all possible estimations and situations into consideration (Elizabeth Mays, 1998). An example of the outcome of such a model is as shown below: Fig: result of risk analysis using rule based notation based on neural networks (Found at: www.roselladb.com) The results shown above depict the risk as high for customer depending on their past claim history. Thus, customers with more number of claims will be depicted as risky and vice versa. The above figures show that the quality of the outcome depends on the effectiveness to which the model has been built by taking into account all the parameters with proper consideration of each of them. thus, if the modeling logic is good enough, the analysis is a simple affair and would involve simple translation of the logic into software code. The software also contains procedures that constantly update the logic with the latest results, thereby following the pattern of neural networks. The outcome of this application is then used to compare the attributes of every application in order to determine the associated risk. Thus, both in terms of implementation as well as with respect to the logic involved, the rule based notation technique is the best among the available techniques for risk prediction. Thus, it can be seen that the risk prediction and analysis is dependent on mathematical, statistical and artificial intelligence based approaches. The decision on the methods to adopt depends on various factors such as the size of the dataset, the processing resources available, the preferred processing time and the accuracy of results of sought. Improvements to the existing approaches are being sought continuously, and are expected to revolutionize the industry in the near future. REFERENCES 1) Douglas L. Reilly (1996), Risk Assessment of Mortgage Applications with a Neural Network System: An Update as the Test Portfolio Ages. Providence: Nestor, Inc. 2) William Mendenhall (2001), Multiple Regression Analysis, Introduction to Probability and Statistics. Boston: Duxbury 3) McLeod (1993), Predicting Credit Risk: A Neural Network Approach, Journal of Retail Banking, vol. 15, No. 3. 4) Miley Wesson Merkhofer (1987), Decision Science and Social Risk Management: A Comparative Evaluation of Cost-benefit Analysis. London: Barnes & noble. 5) Elizabeth Mays (1998), Credit Risk Modeling: Design and Application. London: Barnes & noble. 6) Chaur G. Wu (2004), Modeling Rule-Based Systems with EMF. Eclipse corner article Read More