INTRODUCTION
As the heart of the insurance business, the underwriting function has remained mostly unchanged for nearly 400 years when Lloyd’s of London was a place where ship owners would seek out men of wealth. The two would contractually agree to share the financial risk, in the unlucky event that the ship would be lost at sea (Gibb,1972; Golding & King-Page, 1952).
Today, insurance underwriters perform a similar function on behalf of their respective insurance companies. Underwriters gathering pertinent information and analyze their potential clients to determine whether or not they should underwrite the risk; and if so, what premium they would require for the insurance policy. Insurance companies employ actuaries to help the underwriter in this process by studying past insurance losses and making predictive models for future risks. Using traditional statistical methods, insurance actuaries look for loss-contributing characteristics within the risk (Webb, Harrison et al., 1992). When the actuaries find positive relationships between the policy characteristics and subsequent losses, they create “underwriting guidelines” for the underwriters to follow, when analyzing potential clients (Malecki & Underwriters, 1986).
For hundreds of years, actuaries used pencil and paper to perform their statistical analysis; it was a long time before they had the help of a mechanical adding machine, still longer before they had computers. As recently as 1981, computers were not considered important to the underwriting process. Leading experts in insurance underwriting believed that the human-judgment factor involved in the insurance underwriting process was too complex for any computer to handle as effectively as a human underwriter (Holtom, 1981).
Recent research in the application of technology to the underwriting process has shown that Holtom’s statement may no longer hold true (Kitchens, 2000; Lemaire, 1985; Rose, 1986). The time for computers to take-on an important role in the insurance underwriting judgment process may be here. The author intends to illustrate the applicability of artificial neural networks to the insurance underwriting process.
BACKGROUND
The American Institute for Chartered Property Casualty Underwriters reports that the most common considerations found in automobile underwriting guidelines are: age of operators, age and type of automobile, use of the automobile, driving record, territory, gender, marital status, occupation, personal characteristics of the operator, and physical condition of the vehicle. Traditionally, these comprise the basic variables used in determining the acceptability, classifying, and rating of private passenger automobile insurance policies (Malecki & Underwriters, 1986).
Private passenger automobile insurance is well suited for artificial intelligence applications applied to the underwriting function. There are three primary reasons for this: there is a fixed set of finite data used to make the underwriting decision; policies are highly standardized; and deviations from the standard insurance contract are rare.
In recent years, researchers have considered the application of computers to the process of automobile insurance underwriting. Two studies attempted to predict the acceptability of a given policy from a broad underwriting standpoint (Gaunt, 1972; Rose, 1986). Two other studies considered the possibility of predicting a loss on an individual-policy basis (Lemaire, 1985; Retzlaff-Roberts & Puelz, 1966). Another study focused the relationship between premium and customer retention from year-to-year. One study was designed to predict losses on individual policies using artificial neural networks (Kitchens,2000).
The recent use of artificial neural networks represents what may result in the most accurate application of computers in the underwriting process. Originally developed in the 1940′s, artificial neural networks were designed to replicate and study the thought process of the human brain (Cowan & Sharp, 1988). Early research showed that all processes that can be described with a finite number of symbolic expressions could be represented with a finite number of interconnected neurons (Whitley, Stark weather et al., 1990). Thus, artificial neural networks also provide a means of economic problem solving.
The author believes that for a number of reasons discussed in the following section, artificial neural networks can be successfully applied to the insurance underwriting process in order to reduce the ratio of insurance losses to insurance premiums.
NEURAL NETWORKS FOR INSURANCE UNDERWRITING
Artificial neural networks were first developed in the 1940′s as a mathematical model used to study the human thought process (Cowan & Sharp, 1988). McCulloch and Pitts in 1943 proved that all processes which can be described with a finite number of symbolic expressions can be represented in a network of interconnected neurons (Whitley, Starkweather, & Bogart, 1990). This makes the artificial neural network a mathematical modeling tool in addition to a representation of the human brain.
Using a data set consisting of dependent and independent variables, an artificial neural network can be trained until it converges on an optimal solution for the dependent variable(s). If properly developed, the resulting model will be at least as accurate as traditional statistical models (White, 1989).
The insurance business, as practiced in the United States, has certain characteristics that produce less than optimal financial results. There are five basic reasons that the unique abilities of artificial neural networks can improve the underwriting process:
First, an artificial neural network model will be successful because the inequity of the current rate classification system will allow neural networks the opportunity to more accurately assess the risk level of each and every individual policy holder, rather than a class of policy holders (Wood, Lilly et al., 1984).
Second, an artificial neural network model will produce improved results because current actuarial methods of study will benefit from the broad range of available tools ” such as more recent developments in the field of artificial intelligence (Cummins & Derrig, 1993; Kitchens,2000).
Third, an artificial neural network model will improve the current state of actuarial research. Traditionally, the primary method of research in this field has been to predict the pure premium (the amount of premium required to pay all of the losses in a given class of insured accounts, a.k.a. “relative rates”). In comparison, actual premiums include the pure premium along with other important factors such as profit margin and operating expenses. The traditionally used pure premium models follow an actuarial approach, but not necessarily an underwriting approach. While it is intended to reduce corporate loss ratios, current actuarial research does not take an underwriting approach to the process. A fresh perspective on the problem could produce improved results.
Fourth, an artificial neural network will produce improved results because historically, statistical models used in predicting insurance losses have been able to produce only marginal incremental improvements. Given the current state of technology, the time has come for new insurance actuarial models to take advantage of the available speed and flexibility of artificial neural networks to solve what is clearly a complex problem, which will require extensive training and is likely to involve a complex architecture (Kitchens, 2000).
Fifth, even if the actuarial models are “perfect” (which the author contends they are not), the neural network should be capable of at least matching the current statistical results, if not improving upon them. This is because artificial neural networks comprise a class of nonlinear statistical models whose processing methods are designed to simulate the functioning of the human brain (Hawley, Johnson et al., 1990). The advantage of neural network models over other modeling methods grows with the complexity of the relationship between input and output variables; however, greater complexity of the underlying relationships between variables requires a more complex design (Lee, White et al., 1993). Provided the appropriate network architecture, a neural network output function can accurately approximate any mathematical function (White, 1989). Further, a model can achieve any degree of desired accuracy if the neural network is properly designed (Funahashi, 1989).
NEURAL NETWORK MODELS: DESIGN ISSUES
Automobile accidents occur with a certain degree of randomness, and it is expected that they will be very difficult to predict on an individual-policy basis. Previous research has shown that an underwriter’s ability to predict the actual value of a paid claim is exceedingly difficult, if possible at all (Kitchens, Johnson et al., 2001). However, a successful system needs only to predict the incident (occurrence) of a loss, not the dollar value. In addition, a successful model would not have to predict each and every accident, as long as the predictions that the model makes are accurate. In fact, a new model needs only to outperform any current models, in order to prove its self worthwhile. As an industry rule-of-thumb, the average loss-to-gross-premium ratio is approximately 60%. The rest of the collected premium is used to pay operating expenses and a small profit of approximately 3%. Thus, if a new model could reduce losses by 1%, it would represent a 33% increase in operating profit. If a corresponding decrease in operating expenses such as loss-adjustment expenses is incurred, the operating profit could be increased by as much as 53%. This in itself is not a justification for using artificial neural networks, but it is enough incentive to try nontraditional techniques.
While it is theoretically possible for a computer program to handle the underwriting function, a reliable model has not yet been developed. Aside from the practical development, the human-side must also be considered. It will take time for society, insurance regulators, and the insurance industry to understand and accept a computer-based underwriting model. The development of a computer-based model that might aid the underwriter in the decision making process will be a good first step. An artificial neural network as an underwriter’s tool could be used in several ways ” to confirm an underwriter’s decision; to provide a suggested course of action; or to handle routine policies, allowing the underwriter to spend more time on more complex policies. As an underwriter’s tool, a model should be useful, reliable, and convenient while providing information of value. While there may be many methods of designing such a tool, the author believes that artificial neural networks hold the greatest likelihood of success.
In the development of an artificial neural network model as an underwriter’s tool, several things must be determined: the output required, the input variables, the type of artificial neural network, the architecture of the artificial neural network, and the interpretability of the output.
The underwriter’s decision-making process boils down to two basic decisions. First, he must decide whether to accept or reject the risk. Second, if accepted, a decision as to the premium that will be charged must be made.
Depending on the purpose of the model, the required output may place stringent requirements on the required input. One reason all previous models have had limitations on their applicability has been due to the lack of quantity or quality in the available data sets used to generate the model.
For purposes of insurance loss-prediction modeling, the Genetic Adaptive Neural Network (GANNT) algorithm is an appropriate choice. The GANNT algorithm is designed to overcome difficulties associated with the popular gradient and backpropagation techniques (Dorsey & Mayer, 1994).
The genetic algorithm was first proposed in 1975 (Holland, 1975; Nygard, Ficek et al., 1992). It was shown that the biological evolutionary process could be applied to an artificial mathematical modeling system (Konza, 1992). The concept is based on the theory that an optimization problem can be encoded as a list of concatenated parameters (nodal weights), which are used in the artificial neural network (Whitley, Starkweather et al., 1990). The genetic algorithm works through a process of modeling founded on the biological process by which DNA replicates, reproduces, crosses over, and mutates (Crane, 1950). These procedures are then modeled in a computer-based algorithm to solve complex problems (Nygard, Ficek et al., 1992). The actual operations of the genetic adaptive neural network training algorithm are explained in detail by Dorsey, Johnson, and Mayer (1991).
RESULTS
Recent research has shown that the automobile insurance underwriting process practiced in the United States is lacking in precision. Underwriters are not efficiently utilizing all of the information available to them. The current human-based underwriting process uses only a portion of the available information. In some cases, so much valuable information is overlooked that the remaining unutilized information can be used to make a more accurate accept/reject decision than the initial underwriter made (Kitchens, 2000). With the benefit of the flexibility and adaptability of an artificial neural network, the unutilized information may be used in the future to make more accurate and more precise underwriting decisions.
FUTURE TRENDS
Future research should be focused on two primary areas. First, to be properly trained, an artificial neural network requires data from both the “accepted” and the “unaccepted” policies. Thus, some effort needs to be focused on obtaining information about the policies that are currently being rejected by the underwriter. This is difficult information to obtain because insurance companies have no reason to track losses on policies they previously rejected. But, this data will be valuable in the development of a more complete underwriting model.
Second, future research should go beyond the ac-cept-or-reject decision, and investigate the premium-setting decision. A model capable of second-guessing an underwriter’s accept-or-reject decision might be capable of reducing an insurance company’s losses. But, a model that can both accept-or-reject, and set the premium, might be capable of reducing the cost of underwriting, streamline the business process and produce policies that are more appropriately-priced.
CONCLUSION
Since insurance underwriting began 400 years ago, until recently, the biological neural network (human brain) has been the fastest, most efficient, and most readily available information processing tool available. It has naturally been the tool of choice for underwriters when making accept-or-reject and pricing decisions on insurance policies.
In 1981, it was a common belief that computers could not be used to replace insurance underwriters (Holtom, 1981). In the past 20 years or so, computers and technology have made tremendous advancements. During the same time period, sophisticated mathematical models and the algorithms used to generate them, including the Genetic Adaptive Neural Network Training Algorithm, have taken advantage of increased computing power and availability. Artificial neural networks and the technology used to run them have been shown to out perform the traditional human-based mental decision making practices, in both speed and accuracy; if only for limited domains and applications, such as insurance underwriting.
KEY TERMS
Actuary: A statistician who practices the collection and interpretation of numerical data; especially someone who uses statistics to calculate insurance premiums.
Artificial Neural Network: (commonly referred to as “neural network” or “neural net”) A computer architecture, implemented in either hardware or software, modeled after biological neural networks. Nodes are connected in a manner suggestive of connections between the biological neurons they represent. The resulting network “learns” through directed trial and error. Most neural networks have some sort of “training” algorithm to adjust the weights of connections between nodes on the basis of patterns found in sample or historical data.
Backpropagation: A learning algorithm for modifying a feed-forward neural network which minimizes a continuous “error function” or “objective function”. Back-propagation is a “gradient descent” method of training in that it uses gradient information to modify the network weights to decrease the value of the error function on subsequent tests of the inputs. Other gradient-based methods from numerical analysis can be used to train networks more efficiently.
Biological Neural Network: A network of neurons that function together to perform some function in the body such as thought, decision-making, reflex, sensation, reaction, interpretation, behavior, and so forth.
Genetic Algorithm (GA): A class of algorithms commonly used for training neural networks. The process is modeled after the methods by which biological DNA are combined or mutated to breed new individuals. The crossover technique, whereby DNA reproduces itself by joining portions of each parent’s DNA, is used to simulate a form of genetic-like breeding of alternative solutions. Representing the biological chromosomes found in DNA, genetic algorithms use arrays of data, representing various model solutions. Genetic algorithms are useful for multidimensional optimization problems in which the chromosome can encode the values for connections found in the artificial neural network.
Insurance: Protection against future loss. In exchange for a dollar value (premium), insurance is a promise of reimbursement in the case of loss. Contractual arrangement of insurance may be voluntarily or by government mandate (such as minimum requirements for automobile insurance for licensed drivers).
Nodal Connections: Connections between nodes in an artificial neural network. They are communication channels that carry numeric data. They simulate the axons and dendrites used to carry electrical impulses between neurons in a biological neural network.
Node: A mathematical representation of a biological neuron. Multiple layers of nodes are used in artificial neural networks to form models of biological neural networks.
Risk: An individual or organization’s exposure to a chance of loss or damage.
Underwriter (Insurance Underwriter): An employee of an insurance company whose job duties include analyzing an application for insurance and making a decision whether to accept or reject the application. If accepted, the underwriter further determines the premium to be charged. Underwriters also review existing insurance contracts for renewal.
