Share this post:
IBM has sponsored two virtual roundtable conversations where industry experts share views, examples and tips on the impact of COVID-19 on insurance fraud. If you haven’t seen them, you can catch the replays on demand at COVID-19: Responding to the threat of fraudulent claims and How can the Insurance Fraud Industry stay ahead of Insurance Fraud in the new normal?
During both events, the fraud experts discussed the use of data and analytics to inform how fraud controls can keep up with the changes in the real-life behavior from individuals and organizations. Jay Bobrowsky from the State Compensation Insurance Fund urged the audience to “more than ever, to identify your data, to understand it, and act on it.”
During the roundtables there was little time to discuss relevant analytical techniques in depth. In this article I dive a bit deeper into this topic and cover the following:
- The behavior of individuals and organizations change due to COVID-19.
- These changes are impacting your automated fraud controls.
- You need to understand the changes in your claims data in order to respond to them.
- Six advanced analytics techniques to help you monitor and adjust to the changes occurring in the claims you receive.
Behavior continues to change
““Behavior change is individual and not static.””
As you no doubt are experiencing first-hand, COVID-19 continues to change many aspects of our lives. Some of the ways in which my family and I have changed our behavior:
- We’ve arranged our home to allow four people to work and study there, including upgrades to desk chairs and computer equipment for everyone.
- We sharply reduced driving right from the start (yes, I was one of the many people that had to call roadside assistance to get my car battery going again . . . ), and it continues to be far less.
- I’ve actually started to exercise more regularly, not less.
- We were prepared to spend our summer vacation at home, but with travel restrictions easing we’ve gone back to our original plans to venture out-of-country.
- My daughter was initially delighted that her job in the supermarket was a vital profession, but now is more concerned about shoppers abandoning all attempts to keep a safe distance from her.
This list is specific to our family situation and the choices we make; it will look different for others. This is also a snapshot in time, it will not stay this way exactly – some changes will stick, others will revert, and some new changes will occur based on the evolving health, education and workplace situation. Once again, “change is the only constant.”
Impact on AI and Insurance
““Models trained on normal human behavior are now finding that normal has changed.” Assume your automated fraud controls are affected as well.”
Speaking of such change, a McKinsey report talks about “sweeping shifts in economics and consumer behaviors playing out in the data that feed analytical models. This leads to sizeable model drift as pandemic-related data issues tested the robustness of even well-honed and calibrated models.”
An MIT Technology Review article gives examples where pandemic-induced changes in data are affecting AI models for inventory management, investment recommendations, movie recommendations, credit card fraud, email marketing copy generation, order fulfillment and online advertising. The article concludes that “Machine-learning models trained on normal human behavior are now finding that normal has changed, and some are no longer working as they should.”
In that context, it’s not hard to see how my list above of personal changes has a direct bearing on the insurance companies where we have policies for our home, car, health, travel and work. One example that “normal has changed” for insurers is the ongoing analysis by telematics provider The Floow of the effects of the COVID-19 pandemic on mobility and motor exposure. They report a reduction in driving, and a simultaneous increase in risk level per mile driven. If you (or your technology vendor) use automated controls for fraud detection, you should assume it is similarly affected by the behavioral changes of your policyholders and claims suppliers. It’s great if your provider updates their models, but do you know yourself what those behavioral changes are so you can adjust accordingly? And do you monitor the data yourself to see the inevitable changes to the changes in the months ahead?
Six advanced analytical techniques to monitor changes in claims data
With such large shifts in the underlying real-life behavior it’s not simply a matter of retraining your models on new data in an automated manner. Data science platforms can certainly do that, but if you don’t understand the fundamental changes that are playing out, you’re in for a lot of unpleasant surprises.
Below, I recommend six advanced analytical techniques that will help you understand what’s going on in your data in detail. These techniques should be part of your automated controls anyway and can also be used in a more user-driven discovery mode. Use these techniques to compare claims data before, during and after the pandemic. Although these three time periods are conceptually not difficult, determining the exact date for each will take some consideration. It will depend amongst others on the lines of business and geographies you operate in. For example: travel in Europe was already severely reduced before most official lockdowns went into effect (after weeks of reduced air travel, Flybe in the UK went into administration on March 6th, more than two weeks before the UK lockdown was announced).
““Understand fundamental changes in behavior by comparing claims data before, during and after the pandemic.””
Also, the effects of financial support for companies and individuals (including when the money stops coming ) will differ from one place to the other. If you know what’s going on in your data, you’ll be able to make a good decision about the time periods to compare. Tag each claim with the appropriate label and you’ll be able to apply all the following techniques.
1. Text analytics: What is changing in what your customers and providers say about their claims?
Extract the words used in loss descriptions and other text data sources associated with your claims, then look at how word usage changes between the before, during and after time periods. Is the frequency of words changing? Are new words popping up? Is one word replacing another?
During our Covid-19 virtual roundtable, Matthew Smith from the Coalition Against Insurance Fraud gave the example of scammers going door-to-door offering home remodeling – that behavior is going to show up in text data somewhere, maybe in customer service calls to ask about coverage, maybe in claim descriptions. If normally you see this, say, once a month, and now you get three incidents in a week from the same area, that’s a signal to pick up on.
2. Entity resolution: What is changing in the entities you do business with?
Entity resolution identifies the same person or organization, even if some details are different due to intentional obfuscation or accidental data entry. It is not uncommon to have a reduction in entities of up to 30% through the resolution process.
This is a process that typically occurs in the background, but it is insightful to look at the results of the entity resolution across the before, during and after time periods. Are there entities that have an increasing number of duplicates? Are entities being resolved in an unusual or unexpected way now (e.g., involving employees)? Are people obfuscating their entity to get cover they otherwise wouldn’t?
During the virtual roundtable, Joe Wehrle from the National Insurance Crime Bureau touched upon the topic of identity theft. Tom Gardiner from Aviva emphasized the need to look across Claims and Point of Sale. Both of those areas would benefit from this analytical technique.
3. Network analytics: What is changing in the relationships between parties?
When entity resolution has told you “who is who,” the next step is to find out how parties are connected to each other, i.e., “who knows whom.” Is there a new or increasing concentration of claims around specific suppliers? Is there a change in risk from associations with known fraudsters? Are there unusual or unexpected connections between entities, for example, involving employees?
During the virtual roundtable, Celeste Dodson from the International Association of Special Investigation Units made the differentiators between opportunistic and organized fraudsters. This technique will particularly help to keep track of the latter, for example to spot re-emergence of staged accidents.
4. Claims segmentation: What different types of claims are you now receiving?
As the pandemic evolves, the mix of claims that you receive will be changing and this goes deeper than just “more workers comp claims.” Segmentation will help you understand this more deeply by finding groups and subgroups of similar claims at a very granular level.
Once you understand the subgroups that appear before, during and after Corona, you can compare these different types over time. How quickly are claims segments increasing and decreasing in volume and amounts? To what extent is there “segment migration” where the same claim would be classified differently in another time period? Are there smaller segments of claims that are new?
During the virtual roundtable, Ben Fletcher from the Insurance Fraud Bureau mentioned that fraudsters will be trying to tweak their approaches. Tom Donahue from the Pennsylvania Insurance Fraud Prevention Authority gave the example of an expected increase in “crash and buy” claims. A granular claims segmentation will help to spot those nuances.
5. Anomaly detection: Are the patterns of anomalies shifting?
Anomaly detection actually starts with first capturing normal behavior by creating peer groups of claims that look similar or claims suppliers that act similarly. It then flags claims within each peer group that are different.
As mentioned above, “normal has changed” and this technique is great to provide insights into that. How is the pattern of “normal” shifting between the time periods? What cases are considered outliers now that would not have been before – and vice versa? How is claimant or provider behavior changing? During the virtual roundtable, Wade Wickre from IBM emphasized these aspects.
6. Supervised modeling: Are the patterns around proven frauds or false positives changing?
Supervised models are trained on historical data to differentiate between outcomes, for example to distinguish referred claims that lead to proven frauds from those that lead to false positives. These models are typically then used to score new claims in an automated manner, but this technique is also very useful to gain deep insights into the structural differences between groups of claims.
For instance, how are your proven frauds different before, during and after the pandemic? Are your false positives changing from one time period to the other? During the virtual roundtable, Glen Marr from IBM mentioned how this technique helps with the continued need to reduce false positives.
As outlined above, COVID-19 is driving behavior change which impacts your automated fraud controls. The six advanced analytics techniques help you understand what changes are occurring in order to determine how to adjust your controls.
If you have questions about any of the above, do not hesitate to contact me. Also, I’d love to hear if you’ve already applied these techniques and what value they brought.
Martijn Wiertz, Principal Presales Consultant, IBM Financial Crimes Insight
Martijn Wiertz, IBM