Data mining to stem insurance fraud down under

Operating under a tight government budget in a difficult climate for insurance companies throughout Australia, WorkCover found investing in data-mining tools a “necessary business requirement” in its ongoing effort to cut down on fraud.

WorkCover Authority, which manages workplace injury management and workers compensation systems within New South Wales, established a Compliance Improvement Branch to focus on claimant fraud, under insurance and non-insurance.

To this end, the branch found it essential to build a research team and initiate a data mining project. The aim was twofold: to identify employees making inaccurate insurance premium payments and to cut down on illegitimate insurance claims.

The first step was to establish the Data Mining Analytical Research Team (DMART) and roll out data-mining tools.

Peter O’Hanlon, team leader of data-mining research at WorkCover and an original member of DMART, said the investment in data-mining software was a “necessary business requirement.”

“There were enough savings calculated to justify the expense of going this way. To not go this way would have cost so much more. To have left it as it was would have cost us so much more,” he said.

“This branch had been established and the analytics driving the activity was fundamental. It was never going to be a big branch and the budget was tight. But we had the money to spend to see the potential savings (from this investment).”

O’Hanlon would not comment on the tougher climate for the insurance company, but said, “there has been a greater impetus to see to it people pay us premiums and to cut down on fraud.”

Prior to the data mining tool, investigations into premium avoidance and insurance fraud was done in a reactive manner, according to O’Hanlon. “We developed technology that helped us proactively seek out and detect cases of fraud and underpayments of premium,” he said. “WorkCover didn’t have this focus in the past when looking at both areas.”

In order to home in on inaccurate premium payments, such as the under declaration of wages, the team conducted targeted wage audits. It can take three to 12 months to do an audit and, with approximately 350,000 employees paying premiums to the authority, audits can be a costly process, causing a major inconvenience to employers. With the new system, O’Hanlon and his team can get an employee history and give a score of risk for each employer in the program.

“The system helped to fine tune to predict which audit not to do, that would be a waste of time and money,” he said.

“To run an investigation, we create a data set that contains known cases with our collected historical information. Then, using different decision trees and neural networks, we build a model that predicts behaviour. We validate that it works with previous known cases. Finally, we apply that model to cases that we haven’t looked at. From this we find potential new cases.”

Targeted audits helped the team see a return on its investment.

“We can now do 1,000 wage audits a month, previously we did 300 to 500 a month. So we increased the number and improved the selection process. We doubled the premium collected two years in a row. This is due largely to the implementation of data mining and is part of the ROI (return on investment),” said O’Hanlon. “We can generate those lists monthly of every employers, then we simply take the top 1,000 per month.”

The system also helps the team to pinpoint fraudulent work cover claims. Although there are fewer cases of claimant fraud, compared to underinsured employers, O’Hanlon said those few cases can generate significant losses to the company.

“One particular case of fraud alone saved us A$700,000 (US$389,410) when we caught it. We picked it up as a result of the program,” he said.

Claimant fraud investigations can be very involved and lengthy. O’Hanlon would not give a number of the amount of cases that the team have investigated as a result of the platform, but did say they have screened hundreds of cases for indications of possible fraudulent behaviour.

O’Hanlon was sketchy on the detail involved in fraud investigations.

“We look at claims made by the workers and generate lists of high risk claims on a more ad-hoc basis. Then we screen through databases and use manual searches to do a full investigation,” said O’Hanlon.

“From the outset, we recognized the changing nature of fraud cases over time. So we’re not looked into a system for good, because the nature of fraud will change.”

O’Hanlon could not disclose the investment into the program, but said the authority realized its return on investment within the first three months. The program has been running since October 2001.

“This was measured in actual savings to the scheme, such as the premiums collected,” he said.

“WorkCover has received A$11 (in) return for every one dollar it invested in cracking down on claimant fraud, and A$6.30 for every dollar in wage audit.”

At the moment, only three people are using the data-mining application within WorkCover. However, other business units within WorkCover have picked up on the benefits that the compliance branch has seen and are looking to adopt the application in a strategic way, including the Insurer Performance unit, Scheme Monitoring unit and workplace safety.

It took one and a half years to get the data mining tool to a stage where DMART felt its risk management system was working effectively, and just months for the tool to be running completely, according to O’Hanlon.

The assessment of competitors was done in-house, with not too much vendor involvement. But if O’Hanlon said if he could do it again, he would get vendor involvement.

WorkCover was already using Base SAS for analytics, but according to O’Hanlon, he couldn’t do any sophisticated modelling that his team does with SAS Enterprise Miner now.

“We were already cutting code in it. Also the SAS team had good knowledge and understanding of our business needs and data mining. They provided a full solution rather than a tool how to apply it.”

Related Download
The Landscape of Self Service Analytics Sponsor: IBM
The Landscape of Self Service Analytics
Download this report to examine the current state of self-service analytics across all industries and company sizes, and view the technology decisions and analytical performance of organizations that reported high levels of self-service in their analytical use base.
Register Now