white paper Uncovering More Insurance Fraud with Predictive Analytics Strategies for Improving Results and Reducing Losses April 2012 Summary Predictive analytics are a powerful tool for detecting more insurance claims fraud with greater certainty, helping insurers contain losses while streamlining claims processing and making claims review more productive. How much value analytics deliver, however, depends on how effectively they are used. The challenge is for insurers to know at what point they are getting the desired ROI how much fraud is being uncovered, how much loss avoided, how much time saved. This white paper sets forth a methodology for determining the optimal use of predictive analytics in claims reviews to protect against fraud. www.fico.com Make every decision count TM
Introduction It is virtually impossible to determine exactly how much money is stolen through insurance fraud globally each year because most of it goes undetected. The number of cases that are detected is staggering, but it is quite likely much lower than the number of acts of fraud actually committed. Efforts to determine the true magnitude of the problem, therefore, involve some estimation and guesswork. However, a sampling of different regions and industry segments makes it clear it is a global, industry-wide and growing issue. According to estimates by the Insurance Information Institute, insurance fraud accounts for about 10% of the property/casualty insurance industry s incurred losses and loss adjustment expenses. In the United Kingdom, the Insurance Fraud Bureau estimates losses due to insurance fraud amount to around 1.9 billion (nearly $3 billion), causing a 5% increase in insurance premiums. The Insurance Bureau of Canada estimates that personal injury fraud in Canada costs about C$500 million annually. A 2010 study by the Brazilian Insurance Confederation found that, across all segments except health, suspected claim fraud totaled R$1.9 billion ($1.05 billion), around 9% of the value of all claims studied. According to a 2010 estimate, auto insurance fraud losses in the US alone are approximately $8 billion, and add $200 to $300 a year to individual insurance premiums. The types of fraud range from opportunistic soft fraud, such as adding previous damage to a current claim, to conspiring with doctors or repair shops to inflate bills, to hard frauds like staged accidents, phony injury claims and claiming false hit-and-runs. Across all segments, the insurance industry is constantly working to get ahead and stay ahead of fraudsters. The fraud-fighting arsenal holds a variety of tools. Some of the most powerful tools employ predictive analytics. While these tools are generally effective at identifying fraud, knowing exactly how to use them to optimize the savings-to-workload ratio can be confusing. In this white paper, we lay out a methodology to optimize the use of predictive analytics, and explore how different usage strategies can impact potential savings and recoveries on the fraud front.»» The Auto Insurance Example Among all categories of personal insurance, auto insurance is particularly vulnerable to fraud because it has an unusual number of layers and levels of human intervention. The payouts resulting from an accident include first-party repairs, third-party payments and personal injury compensation. Fraudsters need to look for just one weak link in the long claims process. An act of fraud can involve any of a number of players, working individually or together, such as the driver, repair shops, medical service providers, a fraud investigator or a third party involved in the accident. It is not uncommon for different players to conspire with each other and operate as formidable, organized criminal rings. In recent years, the number of organized gangs operating in insurance fraud has gone up across geographies. Insurance companies have a wealth of data available at multiple levels, including policy, accident, claims and payment-level information. This volume of data provides the opportunity for rigorous investigation to identify patterns in fraudulent claims and look for fraud indicators. While some of the fraud patterns vary from geography to geography, there are a number of patterns that can be found across the globe. Operations of organized crime rings are often localized, making some locations riskier than others. A large number of these types of claims tend to be staged accidents or scavenged vehicles that get reported as theft. Fraudsters often burn the vehicles in these cases after removing parts, then file lossdue-to-fire claims. Most people involved in these claims, including the driver, third-party claimant, witnesses and others, are from the same locality and have been involved in other incidents, often with interchanged roles. 2012 Fair Isaac Corporation. All rights reserved. page 2
Using Predictive Analytics to Identify Fraud Whether in auto, homeowner s, property and casualty, health or any other category, predictive analytics can be a powerful tool to identify fraud in insurance claims. By building a fraud model that rank-orders claims based on their likelihood of fraud, you will be able to focus your review time where it is most likely to result in identified fraud and in savings. The power of an analytic model is that it takes multiple features into account simultaneously, and uses them to identify cases with a very high likelihood of fraud. For example, an unusually high number of fraudulent accidents or incidents occur very close to the date of inception or expiration of the policy. However, not every claim filed near the inception or expiration date is fraud. Taking these features into account in an analytic model in combination with multiple dimensions simultaneously can provide a very good and accurate indication of the likelihood of fraud. The ability to rank-order claims by likelihood of fraud is a powerful triaging technique. But how do you know how many claims to work? How do you determine if a model is working well enough to incorporate into your claims processing workflow? Getting More Hits An analytic model is never perfect. Claims with a higher likelihood of fraud will be pushed to the top score bands, but even in the very highest score bands there will be a number of non-fraudulent claims. The idea behind an analytic model is to increase the likelihood of finding fraud in a set of claims under review. For instance, if reviewers just look at claims at random, they may find one in every 20 claims to be fraudulent. By using an analytic model and focusing on the claims identified as most likely to be fraudulent, that rate may be increased substantially, perhaps to as many as one in four. Essentially, the analytic model increases the richness of the pool under review. FIGURE 1: TYPICAL HIT RATE BY SCORE 1000 800 600 Score More actual fraud is likely to be found among claims that score highest for likelihood of fraud. 400 200 0 30% 25% 20% 15% 10% 5% 0% Hit rate A useful way to measure this increased richness is through a metric known as the hit rate. The hit rate provides a measure of performance that is easy to compute and intuitively appealing. It is defined as the number of true positives (definitely fraudulent claims) out of the total number of claims reviewed. If you review 25 claims and find 5 of them to be fraudulent, that would translate to a 20% hit rate. Because a predictive model produces a rank-ordered list, you would expect the hit rate to decrease as you work down the list. The highest hit rate should occur among the claims most likely to be fraudulent, which are the ones at the top of the list. A typical distribution of the hit rate could look something like that shown in Figure 1. In this case, you begin working the highest scoring cases and find that you have a hit rate in the 25% range one out of every four claims a reviewer looks at will turn out to be fraudulent. As you work cases lower in the score range, the hit rate decreases. 2012 Fair Isaac Corporation. All rights reserved. page 3
FIGURE 2: IMPROVEMENT IN DETECTION USING AN ANALYTIC MODEL Gains Plot Hit rate 100 90 80 70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 Cum, % Target = 0/-1 Cum, % Target = 1 KS = 62.5 80 90 Cumulative Percent of Population by Model Score 100 Now, how do you determine if the model is performing up to a reasonable set of quality standards? How do you determine where a good operational threshold lies? Though the hit rate is an intuitive measure, it makes it difficult to answer these questions. Another way to pose the question is to ask how many cases you have to work to identify an acceptable level of fraud that you think is occurring in your system. An analytic model will not be able to identify all of the fraud in a claims system in the highest score distributions. But a good measure of model performance is quantifying how much of the fraud you can identify within certain score bands. This is best represented in a gains chart (Figure 2). Without an analytic model, if we just reviewed claims at random, we would expect to find a level of fraud that matched the number of claims we reviewed. For example, if we had 100 claims and reviewed 20% (or 20) at random, we would expect to identify 20% of the total fraudulent claims in that review. If the overall fraud rate was 10%, that means we would expect to find two fraudulent claims in the reviewed batch of 20. That level of detection is represented by the orange line in the gains plot in Figure 2. Without any review strategy or triaging, it s what we would expect. It s the worst we could reasonably expect to do in terms of fraud detection. The yellow line shows an example of the gains in detection you can realize with an analytic model. In this example, if we reviewed 10% of the top scoring claims (a point represented by 10 on the horizontal axis), we would identify 70% of the total fraud. One measure of model performance is to identify the biggest discrepancy between the orange know nothing line and the yellow analytic model line. In this case, that maximum separation comes at about 16% of the claims capturing 78.5% of the fraud. The difference between these two numbers is a statistic known as the KS, which is often used to measure model performance. This type of graph is useful in determining how much known fraud is captured by the model. It provides better guidance than the hit rate for setting an operational threshold, but it s still not clear exactly how many claims should be worked. Setting Operational Levels When evaluating model performance, there are two factors that should be considered. Does the model generate a pool of claims with enough fraud that it s worth a reviewer s time to manually review the claims? And how big is the set of claims that is worthy of review? Assumptions In order to determine if it s worth a reviewer s time to review claims at a given score threshold, you need to know how many fraudulent claims a reviewer can identify in a given time frame, the cost of identifying those fraudulent claims, and the savings anticipated from their identification. For example, let s say a reviewer can identify one fraudulent claim in an hour. If that fraudulent claim pays, on average, $3,000, then the reviewer has identified $3,000 in fraud in an hour. If a reviewer can stop payment from going out on that claim, that s $3,000 in savings realized. If some of the amount on that claim has already been paid out, then the savings will be some fraction of that $3,000. As long 2012 Fair Isaac Corporation. All rights reserved. page 4
as the reviewer s time costs less than the amount of savings realized, it makes sense to review claims at that hit rate level. R < h(p f + rp p ) * n r where R = hourly reviewer cost h = hit rate p f = average per claim future payout p p = average per claim already paid out r = recovery rate n r = number of claims reviewed per hour Using this simple formula, it is straightforward to figure out where a reasonable operational threshold lies. Assuming everything in the function other than hit rate is constant across score bands, the savings per hour of review will monotonically decline as the hit FIGURE 3: SETTING OPERATIONAL THRESHOLDS rate by score band declines. Figure 3 shows an example of savings Savings per hour 600 500 400 300 200 100 realized per hour using the above formula. The orange line indicates the per hour cost of review, in this case assumed to be $100. It is clearly well worth a reviewer s time to review the very highest scoring claims. At this level, a $100 investment in review time will net savings of over $500. Where the savings per hour line crosses the reviewer cost line is the optimal place to set the operational threshold in this case, around a score of 700. Working any claims below this score will not result in substantial enough savings to justify the reviewer s time. Conversely, not working all of the claims that score above 700 0% would leave some potential savings on the table. By not working 1000 800 600 400 200 0 Score the full set of claims scoring above the threshold, the analytic model is not being utilized to its fullest extent. Assuming $100 per hour in review costs, reviewing claims in the highest score Once this operational threshold is set, the full value of the model bands can produce significant savings. is relatively straightforward to compute. The model will produce a score for each claim, and it is easy to determine which claims score above the operational threshold. Using this information, in conjunction with the hit rate at each score band, you can quickly determine the potential savings that the model can generate. This leads to a valuation of the power of the model. There is some cost to realizing this value, as reviewers need to work the selected claims. This can all be quantified. t s = h i (p f + rp p )n i (cn i /r) i=1 Dollars per Hour where s = net savings t = number of score bands above operational threshold h i = hit rate in score band i p f = average per claim future payout p p = average per claim already paid out r = recovery rate n = number of claims in score band i c = hourly review cost r = claims reviewed per hour 2012 Fair Isaac Corporation. All rights reserved. page 5
Conclusion Fraud is pervasive in the insurance world around the globe. One of the most powerful ways to detect it is to use an analytic model. Optimizing the use of the analytic model can prove to be tricky. The value of an analytic model lies in the set of high scoring claims produced by that model that are likely to be fraudulent. How large that set is depends on where the operational threshold for the model is set. Setting the operational threshold is an empirical exercise. You can start with some initial assumptions and establish a conservative operational threshold. As you gain experience with the analytic model, you can refine the operational threshold, and modify the size of the review team proportionately to handle the workload. Computing the total value of the model is dependent on where the operational threshold is set. Again, this can be estimated before going into production using some estimations of each of the parameters. Those estimations can be refined as data on model performance and score distributions in operation are gathered. Ultimately, insurers who go through the exercise of determining an effective operational threshold and optimizing model performance stand to realize significant ROI in the form of lower fraud losses, as well as operational and labor savings in fraud detection. About FICO FICO (NYSE:FICO) transforms business by making every decision count. FICO s Decision Management solutions combine trusted advice, world-class analytics and innovative applications to give organizations the power to automate, improve and connect decisions across their business. Clients in 80 countries work with FICO to increase customer loyalty and profitability, cut fraud losses, manage credit risk, meet regulatory and competitive demands, and rapidly build market share. FICO also helps millions of individuals manage their credit health through the www.myfico.com website. Learn more about FICO at www.fico.com. For more information North America toll-free International email web +1 888 342 6336 +44 (0) 207 940 8718 info@fico.com www.fico.com FICO and Make every decision count are trademarks or registered trademarks of Fair Isaac Corporation in the United States and in other countries. Other product and company names herein may be trademarks of their respective owners. 2012 Fair Isaac Corporation. All rights reserved. 2857WP 04/12 PDF