O.R. ApplicationsStrategies for detecting fraudulent claims in the automobile insurance industry
Introduction
Fraud has become a high priority for insurers. For the European insurance industry the Comité Européen des Assurances (1996) estimates that the cost of fraud is over 2% of the total annual premium income for all lines of business combined. In most European countries claim fraud is estimated to represent between 5 and 10% of the total yearly amount of indemnities paid for non-life insurance. In the United States, the Coalition Against Insurance Fraud (2001) states that more than 6% of each insurance premium goes to fraud. The Insurance Information Institute (2004) estimated property and casualty (P&C) claim fraud at $31 billion in 2002.
If not properly addressed, insurance fraud not only puts the profitability of the insurer at risk, but also negatively affects its value chain, the insurance industry, and may be extremely detrimental to established social and economic structures. Moreover, all honest policyholders are victims. Fraud is widely believed to increase the cost of insurance. This cost component is borne directly by all insured parties in the form of increased premium rates. In the end, fraud represents a threat to the very principle of solidarity that keeps the concept of insurance alive (Guillén, 2004, Viaene and Dedene, 2004).
Economic theory has studied P&C claim fraud in depth. Picard (2000) provides a good overview of this literature. It can be shown, for example, that an insurance firm needs to commit to a claim audit strategy to ensure solvency. It has also been shown that as long as there is a good chance of not being caught, an individual filing a claim has a clear economic incentive to defraud.
The most effective way to fight fraud for an insurer is, of course, to prevent abuse of the system. Yet, fraudsters always seem to find new ways of exploiting the inertia of complex systems, especially when a lot of money is involved. It is then imperative to ensure that fraudulent activity is identified at the earliest possible moment, and that persons cheating the system are swiftly tracked down. Since fraud is not a self-revealing phenomenon, insurers typically have to commit considerable resources to its detection. The process of in-depth investigation of suspicious claims is known as costly state verification, because the true nature of a claim (fraudulent or not) can only be discovered by means of an in-depth investigation (see e.g. Bond and Crocker, 1997, Crocker and Tennyson, 1999, Boyer, 1999, Picard, 1996, Picard, 2000). It has been shown that such claim auditing has both a detection and deterrence effect (Tennyson and Salsas-Forn, 2002).
The problem in detecting fraudulent claims is the identification of the characteristics that distinguish them from valid claims. Most insurers train their front-line claims adjusters, often with the help of state- or country-level fraud bureaus, to recognise claims that have combinations of features that experience has shown to be typically associated with fraudulent claims. In many practical situations, though, the identification of suspicion during claims handling continues to be rather subjective in nature. Many insurers still leave it essentially up to the individual adjuster to somehow put together potential warning signs into an aggregated assessment of suspicion of claim fraud. Moreover, since customer service has to a large degree become synonymous with processing efficiency in the context of claims handling, adjusters have no natural incentive to actively look for warning signs. In today’s configuration there is little time for running through extensive lists of fraud indicators at claim time. In other words, one of the core challenges for contemporary fraud detection is to identify fraud in an automated, high-volume, online transaction processing environment without jeopardising the advantages of automation in terms of efficiency, timeliness and customer service.
Some P&C insurers use automated detection systems to help decide on whether to investigate claims suspected of fraud. Automated types of fraud detection should make it possible to reduce the lead-time for fraud control and allow for more optimal allocation of scarce investigative resources. The embedded detection models typically consist of scoring devices relating fraud indicators to some measure of suspicion of fraud. This is where insurers may build on considerable past investments in more systematic electronic collection, organization and access to coherent insurance data. This, among other things, enables the use of algorithmic pattern recognition techniques to create models that help with the identification of insurance fraud.
In this paper we deal with algorithmic learning to score claims for suspicion of fraud. Specifically, we focus our attention on screening activity early on during the life cycle of a claim. What we see in practice is that during the construction phase these scoring models often focus on minimising error rate rather than on cost of classification (see, Derrig, 2002). In this paper we show that focusing on cost rather than error of classification is a profitable approach. This question was already pointed to by Dionne et al. (2003), who work with an average cost approach. In relation to the costs of fraud detection we take into account information on damages and audit costs available in the early part of the claim screening process. In an empirical experiment we discuss several scenarios, the effects of which are explored using real-life data. The data set that is used for this exploration consists of automobile claims closed in Spain that were investigated for fraud by domain experts and for which we have detailed cost information. Thus, the focus of this paper will be on vehicle damage claims in the context of automobile insurance. The findings suggest that with claim amount information available early on in the screening process detection rules can be accommodated to increase expected profits. Our results show the real value of cost-sensitive claim fraud screening and provide guidance on how to operationalise this strategy.
The rest of this paper is organized as follows. Section 2 highlights the main steps in the implementation of a claim fraud detection strategy for a typical P&C insurer. Section 3 tackles the mechanics of a cost-sensitive classification, i.e. the methodology used to build learning programs for fraud detection that reduce the cost of classification rather than the error rate. Section 4 sets the stage for our complementary case study. It covers the data characteristics for the set of Spanish damage claims used in the empirical part and takes a look at the economics of fraud detection for this case. Section 5 projects the mechanism of cost-sensitive classification onto the claim classification setting at hand. This section contrasts six alternative cost incorporation scenarios based on different assumptions concerning the available cost information for classification early on in the process of claim screening. In Section 6 we synthesise our conclusions from the discussion in the previous sections.
Section snippets
Fraud control for the P&C insurer
The generic operational claim fraud control model for insurers (see Fig. 1) includes screening, investigation and negotiation/litigation phases. It is embedded in the insurer’s claims handling process. Claims handling refers to the process that starts with a claim occurrence and a report from the policyholder and ends with the payment, or denial of payment for damages covered. Fraud, in principle, is the only reason for denying payment for covered damages. Fraud, primarily a legal term,
Cost-sensitive decision making
Statistical modelling and data-driven analysis allow for the modernization of the fraud detection process with sophisticated, (semi-)automated, intelligent tools such as unsupervised and supervised pattern learning. Classification is one of the foremost supervised learning tasks. Classification techniques are aimed at algorithmically learning to allocate data objects, described as predictor vectors, to pre-defined object classes based on a training set of data objects with known class labels.
Data set
We have a random sample of claims from a large Spanish insurer for car damages from accidents that occurred in Spain during the year 2000. All the claims included here were audited and the insurer classified them in two categories, i.e. honest or fraudulent, after the investigation process. The data set contains 2403 claims, of which: 2229 are legitimate and 174 fraudulent. This means that about 7.24% of the claims in our data set are fraudulent. There does not seem to be unanimous agreement
Results and discussion
Six possible scenarios are studied. The scenarios differ as to assumptions concerning the available cost information early on in the claim screening process. The presentation sequence of scenarios follows a natural progression. We start with a first benchmark Scenario 1 in which we assume that no cost information is available to the insurance company at screening time. Scenario 2 models the other extreme, in which all claim-specific cost information is assumed to be known at the time of
Conclusions
Many P&C insurance companies are looking for new strategies to tackle claim fraud. Some companies have already implemented automated fraud screening systems based on analysis of characteristics of such interconnected business objects as claims, insured parties, policies and vehicles. These systems, often based on traditional quantitative techniques such as logistic regression and linear or quadratic discriminant analysis, will ultimately be evaluated in terms of profitability. Few, however,
Acknowledgement
We thank the Spanish Ministerio de Educación y Ciencia FEDER grants SEJ2004-05052 and SEJ2005-00741.
References (36)
- et al.
Modelling different types of automobile insurance fraud behaviour in the Spanish market
Insurance: Mathematics and Economics
(1999) - et al.
Hardball and the soft touch: The economics of optimal insurance contracts with costly state verification and endogenous monitoring costs
Journal of Public Economics
(1997) Auditing claims in the insurance market with fraud: The credibility issue
Journal of Public Economics
(1996)- et al.
Detection of automobile insurance fraud with discrete choice models and misclassified claims
Journal of Risk and Insurance
(2002) - Belhadji, E.B., Dionne, G., Tarkhani, F. 2000. A model for the detection of insurance fraud. Geneva Papers on Risk and...
When is the proportion of criminal elements irrelevant? A study of insurance fraud when insurers cannot commit
- et al.
Using Kohonen’s self-organizing feature map to uncover automobile bodily injury claims fraud
Journal of Risk and Insurance
(1998) - et al.
Fraud classification using principal component analysis of RIDITs
Journal of Risk and Insurance
(2002) Insurance fraud
British Journal of Criminology
(1989)- Coalition Against Insurance Fraud, 2001. Annual report. Washington,...
Costly state falsification or verification? Theory and evidence from bodily injury liability claims
Insurance fraud and the Monday effect in workers compensation insurance
Assurances
Insurance fraud
Journal of Risk and Insurance
First steps to fight workers’ compensation fraud
Journal of Insurance Regulation
Fuzzy techniques of pattern recognition in risk and claim classification
Journal of Risk and Insurance
Moral hazard, optimal auditing and workers’ compensation
Cited by (68)
Insuring future automobility: A qualitative discussion of British and Dutch car insurer's responses to connected and automated vehicles
2022, Research in Transportation Business and ManagementThe value of cross-data set analysis for automobile insurance fraud detection
2022, Research in International Business and FinanceCitation Excerpt :The results indicate how to identify misclassified claims. Pinquet et al. (2007) and Viaene et al. (2007) both examined insurance fraud using a real database of Spanish automobile insurance claims. Pinquet et al. (2007) used a two-equation model on audits and fraud, and a bivariate Probit model with censoring was employed on a sample of suspicious claims for which adjusters were allowed to make the audit decision.
Blockchain in construction management: Applications, advantages and limitations
2022, Automation in ConstructionCitation Excerpt :Finding the responsible for a delay Audit technologies can manage disputes since they help managers detect claim fraud during the claim screening process [66]. Since a smart contract provides a system that can track the progress status and manage financial transactions, the responsible persons for any delay can be defined.
Car insurance fraud: The role of vehicle repair workshops
2021, International Journal of Law, Crime and JusticeAn evolutionary approach to fraud management
2020, European Journal of Operational Research