Fighting Fraud with Advanced Analytics

Every single homeowner and vehicle owner is a victim of insurance fraud because the costs associated with fraud translate into higher premiums. Like a hidden tax, fraud losses are factored into financial models so we all pay for it in the end. Fraud costs the global insurance industry billions of dollars every year. Besides the actual losses that occur from fraudulent claim payments, costs include the investments made in resources such as analysts, investigators, prosecutors and technology to detect fraudulent activity. Also costly is the wasted effort — not to mention the negative effect on the consumer — of taking action against a suspected account or transaction that ultimately proves to be legitimate.

Industry estimates vary from country to country, but are consistent in that fraud is a noticeable cost and a growing concern. The estimated cost for general insurance fraud in Canada is Cdn$1.3 billion per year. This is equivalent to saying that approximately 10-15% of the claims paid out annually are fraudulent.

So why are insurance companies not doing more to detect and prevent fraud? The usual answer is that fraud detection is not “cost effective” and, if it is done incorrectly, can have a negative impact on customer service levels. A 2007 survey of North American property and casualty, life and health care insurers1 noted the top three challenges to fighting fraud were the excessive cost of investigation, delayed claims adjudication and that it’s ultimately cheaper to pay the claim.

Historically, fighting fraud has been expensive and inefficient because it is an extremely labor-intensive activity. Today, however, advanced analytics technology can be used to make fraud-fighting processes more effective, efficient and economical. This new technology is able to handle the complexity and massive volumes of information that must be analyzed to identify suspected fraud.

Today’s cutting-edge software tools and techniques provide insurers with the ability to manage the collection, organization, integration and storage of customer and transaction data. Once data is gathered, stored and organized, the application of predictive analytics and other advanced techniques helps identify fraudulent activity. The end result is reduced expenses, increased profits, increased shareholder return, less impact on daily business activities and a more positive image for the insurer.

Taken individually, these techniques and tools are effective on their own in fighting fraud. But taken together in an integrated framework, they are optimally effective and increase the returns to the user.

Insurance Rules and Database Searching

The most common fraud-detection techniques used today employ business rules and comparisons with information from other databases to identify suspected fraudulent activities. Each transaction is tested against a series of predetermined algorithms or rules. Depending on the results of each test (or combination of tests), aggregate scores are assigned and compared to pre-established threshold values to determine the highest propensity of fraudulent activity to take action against (for example high-fail, medium-pass but monitor, low-pass) Such rules can include:

• XclaimsinthelastXyears

• Claim amount over $X, 000

• Policy within XX days of inception/endorsement

• Consults claims chaser (loss assessor) before doctor

• Car “torched” and insured has fire coverage

• Readily admits liability

• Absence of forcible/violent entry

• Registration document in another name

• No police report

• Property damage/personal injury inconsistent

• Multiple versions of accident

• Valuations, receipts are excessive

• X similar claims paid in XX months

• Renewal date less than XX working days

• Changes in sum insured/coverage in last X months

• Whiplash injury < 72 hours after accident (soft tissue damage takes this long, typically, to be felt)

• Time delay in reporting claim

• Insured overly aggressive

• Vehicle struck by rental car

• No independent witnesses -family and friends only

• Inconsistency with police report

• No proof of ownership

• Non disclosure of previous claims or convictions

• Valuations, receipts do not exist

Searching internal and third-party databases is another technique used by many organizations. Firms search these databases looking for matches with existing or prospective customers to determine if the customer is a known fraudster or on a “hot list.”

These are good techniques and used today across the industry. To be effective, however, the rules and databases must be maintained constantly so that they keep up with current fraud activities and schemes, which are ever-evolving and dynamic in nature. What limits the effectiveness of insurance rules and database searching is that they are not predictive and tend to look only at a single customer in isolation.

Anomaly Detection

Anomaly detection is a grouping of techniques that are used to look for outliers or anomalies in data for new or previously unknown fraud patterns.

Profiling is one of the analytical approaches used. When profiling, the analytics tools use models of expected behaviour based on historical data related to individuals or peer groups and then compares that data against current data to find deviations from predicted norms.

Clustering is another technique. In this instance, data is analyzed to identify abnormal groups of claims. Claims that differentiate in relation to a grouping selected via segmentation or profiling are identified for possible further investigation. This technique also identifies where values of outliers are abnormal in relation to each other.

Anomaly detection is beneficial, yet it requires a strong understanding of the data and must be “trained” continually through the building of profiles and seg-ments. Like rules and database searches, it looks at customers or transaction in isolation and has a short life cycle given the ever-changing and advancing nature of fraudsters and their schemes.

Advanced Analytics (Complex Patterns)

Advanced analytics enable companies to generalize fraud patterns for automatic detection by identifying typical patterns and using them to score new applications or claims automatically for fraud propensity. Predictive modelling techniques such as decision trees, neural networks and regression analysis take data associated with historically known fraudulent cases and use advanced analytic techniques to develop models from the data. When applied against new data, these models predict the propensity for a fraud (sometimes known as a “fraud risk score”).

These analytic techniques not only help detect fraudulent activity that has already occurred, they also can be used proactively to determine the propensity of a claimant towards fraudulent activity in the future. The challenge here is that like with the first two techniques outlined above, the evaluation looks at each customer and transaction in isolation.

Social Network Analysis (Associate Link Patterns)

Social Network Analysis (SNA) facilitates detection of unexplained relationships using direct and indirect links between seemingly unrelated parties and transactions. Using modelling techniques such as association/sequence analysis, link analysis/path analysis and fuzzy matching, SNA enables companies to visualize a customer or transaction in relation to other parties and transactions.

SNA uses readily available data such as names, addresses, phone numbers and parties listed on a claim to determine relationships. These relationships can then be reviewed for their propensity to infer fraudulent activity.

An example of this could be as simple as a customer or transaction that is associated with a known fraudster. Not that the known fraudster is the customer or party to the transaction, but rather that they are somehow associated with the customer and transaction — perhaps listed as a witness or an employer. Depending on the associate link, further investigation may be required to determine the propensity for fraud.

Link analysis can also identify “rings” of association — a pattern that shows several parties being linked through the concept of “six degrees of separation” rather than directly. This ring may be coincidence or it may indicate organized crime activity.

SNA has the advantage of looking at information beyond what is associated with a specific customer or transaction. It provides the techniques and visualization capabilities to take into account information associated with the customer or transaction both directly and indirectly, providing a more robust base of information. This enables better fraud detection both proactively and after it has occurred.

A FRAUD FRAMEWORK

Each of the techniques outlined above can be used singly to detect and prevent fraud. However, by far the biggest return on investment in the fight against fraud comes to those firms that integrate the above techniques into a fraud framework. Such a framework enables firms to leverage their investments in data integration activities, case management tools and workforce. The integration of analytic techniques allows for fraud detection once it has occurred and also allows insurers to prevent fraud either at the on-boarding process with the customer or at the time the claim transaction is initiated. A framework integrates analytics techniques that help identify the propensity of fraudulent activity using not only information associated directly with the customer or transaction, but also information associated through links with the customer.

Fraud is a dynamic, fast-moving crime. New schemes arise frequently and fraudsters are becoming more and more sophisticated in their use of technology. As a result, fraud detection and prevention techniques and tools must continue to evolve.

Initially just a rules-based process, fraud detection and prevention now incorporates advanced analytics and social network analysis, driving significant incremental improvements. Yet this is only the beginning. As fraudsters evolve their criminal activities, fraud fighters are expanding their use of current technologies; at the same time, they are preparing to add to their arsenal of prevention tools and techniques within comprehensive fraud frameworks. These next-generation technologies include text-mining software to identify key words and patterns in documents and voice patterning to analyse vocabulary, phraseology, voice tones and inflections.

A successful fight against fraud must be undertaken bearing in mind the growing need to reduce overall fraud-related costs. Today’s advancements in database integration and advanced analytics provide the necessary tools and techniques to accomplish this effectively, efficiently and economically.

1 The State of Claims Fraud Detection and Prevention Survey, CMP research, May 2007.

———

Advanced analytics enable companies to generalize fraud patterns for automatic detection by identifying typical patterns.