Future TenseThe first step in improving profitability is to understand the costs and benefits of current business decisions. Data mining can help.
By J.T. Lehman Continued from Page 1 Two models are necessary because selecting a claim for investigation and the outcome of the investigation are two separate events. The model that predicts which claims will be selected for investigation is built on all claims; the model that predicts which claim investigations will have a good outcome is built on only those claims for which the outcome is known - that is, claims for which an investigation was started. In deployment, the two models are combined. Applying the investigation outcome model directly to all claims would be inappropriate, because claims that have never been investigated qualitatively differ from claims that have been investigated. This difference is problematic. For example, in the U.S. justice system, indictment and conviction are two separate events. Indictment allows a low standard of evidence to prove motive and opportunity. Conviction assumes indictment and a higher standard of evidence. It would be inappropriate to skip the first step of indictment and try everyone for the crime. Similarly, it is inappropriate to skip the step of selection for investigation and proceed directly to estimating how likely an investigation is to be successful. ROI AND RECOMMENDED BUSINESS STRATEGYAs we know, ROI is the gain divided by the cost, where gain is income minus cost. Specifically, in our case, gain is the claim level probability of a good outcome multiplied by the fixed average value of a good outcome minus the fixed cost. Thus, ROI is claim-level probability of a good outcome multiplied by the fixed average value of a good outcome divided by fixed average cost of a claim investigation minus one. For example, if a claim has an 80 percent chance of a good outcome, the average value of a good outcome is $750, and the fixed cost of an investigation is $500, then the gain from the investigation is $100. Conversely, if a claim has a 30 percent chance of a good outcome, then the investigation loses $275. The only element that is variable is the claim-level probability of a good outcome - the output of the predictive model from data mining. Thus, data mining can help estimate the ROI of a claim investigation. The important thing to realize is that claims with a high probability of investigation success have a high ROI, and claims with a low probability of investigation success have a low or negative ROI. Consequently, two business strategies are now available to us: to perform those investigations that have a high expected ROI, and to stop those investigations that have a low expected ROI. For example, if you want a final screen of claims, claims initially selected by adjusters for claim investigation are run through this secondary screen before the claim investigations are conducted. Investigations on claims with a high likelihood of a good outcome proceed; investigations on claims with a low likelihood of a good outcome are routed for review by supervisors or are simply cancelled. The success of this system can be demonstrated by using a real-world test in which both options are implemented, and the results are compared. If a review of the claim investigations practice determines that opportunities for cost containment are being missed, then you can reroute approved claims with a high estimated ROI of investigation back to the investigators. If a review of the claim investigations practice determines that claim investigations is a money-losing operation, then you can stop those claim investigations that have a low estimated ROI. If an investigative unit wishes to keep the number of claim investigations in the future approximately equal to the number of claim investigations in the past, combining these strategies can achieve that goal. Combining these strategies so that the number of additional claim investigations due to the initial screen and the number of stopped claim investigations due to the final screen are equal maintains the total number of claim investigations. Maintaining the total number of claim investigations has the added benefit of maintaining the enforcement effect, the discouragement of spurious or inflated claims due to the threat of investigation. More general business benefits are apparent, as well. First, driving down cost allows more competitive prices, which increases market share, which drives up stock price. Second, executing more successful claim investigations and fewer unsuccessful claim investigations improves relations with customers: Fewer honest customers are annoyed by investigations that never should have occurred. HAPPY FUTURESIn this article, I've explained how data mining can improve the ROI of a business process. Starting with costs and benefits, only one unknown future event - the probability of a good outcome investigation - was necessary in order to estimate the profit or loss of the decision to investigate prior to actually investigating. I derived two business strategies: one to make money by performing cost-saving investigations that weren't planned, and the other to save money by stopping planned investigations that will probably lose money. There is a useful maxim to summarize this article that serves as a short take-away: "If I only knew x, which happens in the future, then I'd take business action y now, and make tons of money." In this case, "If I only knew whether an investigation would be successful, then I'd only do the investigations that will be successful, and I'd save tons of money by doing the right investigations." This maxim could serve as a one-sentence guide to all predictive data mining projects. If you can fill in the x and the y for your business, then you have the opportunity to start a business-driven data mining project. I intend to compile my work powering business strategies with data mining into a book. If you would like me to cover a particular application or have other feedback, please send it to jtlehman@alumni.utexas.net. J.T. Lehman [jtlehman@alumni.utexas.net] creates efficient business strategies leveraging data mining and statistical analysis for Intelligent Technologies Corp. RESOURCESACM Special Interest Group on Knowledge Discovery and Data Mining: www.acm.org/sigs/sigkdd/join.html Kdnuggets: www.kdnuggets.com Berry, Michael J.A., and Gordon Linoff. Data Mining Techniques: For Marketing, Sales, and Customer Support. John Wiley & Sons, 1997 Pyle, Dorian. Data Preparation for Data Mining. Morgan-Kaufmann, March 1999 Related Articles at IntelligentEnterprise.com: "Matching Patterns," April 16, 2001: www.intelligententerprise.com/010416/feat3_1.jhtml "A Meeting of Minds," Nov. 10, 2000: www.intelligententerprise.com/001110/decision1_1.jhtml "Bringing Them Back," July 17, 2000: www.intelligententerprise.com/000717/feat2.jhtml
|
 |
Most Popular This Week
IE Weekly Newsletter
Subscribe to the newsletter
| |
| ||||||||||||||||||||||||||||||||
