In anticipation of his upcoming conference presentation, Predicting Extreme Behavior to improve the rating structure for Travel Insurance, at Predictive Analytics World San Francisco, March 29-April 2, 2015, we asked Richard Boire, Founding Partner of Boire Filler Group a few questions about his work in predictive analytics.
Q: In your work with predictive analytics, what behavior do your models predict?
A: Before even building a predictive analytics solution, we work with the business stakeholders to identify the exact business problem or challenge.
Once this has been identified, we then determine the behaviors that need to be optimized or minimized to solve the challenge/problem.
Our experience and breadth of work has resulted in the following type of model being built:
- Credit risk
- Claim risk
Q: How does predictive analytics deliver value at your organization? What is one specific way in which it actively drives decisions?
A: In building marketing response type models, the ROI is significantly improved as lower costs are achieved to attain the same level of revenue that would otherwise be attained without modelling.
Marketing decisions can then be made on which customers to select based on their ROI.
In building automobile claim risk predictive analytics solutions, we are predicting the loss cost of a given vehicle. By integrating this information with premium, we can then determine which vehicles are overpriced vs. those which are underpriced.
Q: Can you describe a successful result, such as the predictive lift of your model or the ROI of an analytics initiative?
A: Listed below is the case of a property risk model that we built for an insurance company.
The green line represents our predicted analytics solutions while the red line represents the current premium pricing practice of the insurance company. The company’s current pricing practices are delivering value as seen by observing the area under the curve between the red line and the straight line. Yet, predictive analytics solutions built by our company are delivering even further value as depicted by the area under the curve between the green line and the redline.
Q: What surprising discovery have you unearthed in your data?
A: In some cases, we have been able to use the data to identify a major change in someone’s life. A good example of this is identifying an individual that has just moved by observing that the address data has now changed for that individual. The ability to identify movers vs. non movers has led marketers to develop unique communication strategies towards this segment.
Another good example is tenure. We often find that customer behavior is often U-shaped with this variable which implies that newer and longer-tenured customers tend to be loyal, while the middle group or more average tenure type customers tend to be less loyal.
Q: Sneak preview: Please tell us a take-away that you will provide during your talk at Predictive Analytics World.
A: In building predictive risk solutions to estimate the loss amount for this travel insurance company, we encounter huge outliers that often distort or mitigate the overall impact of the model. In this particular case study, we examine an approach that allowed us to in effect produce a better model thereby reducing the impact of outlier events. The key takeaway was that we used a very pragmatic approach in solving the problem. This pragmatism enabled us to combine our domain knowledge of the business along with our data science knowledge which allowed us to arrive at a better overall solution.
Q: What has been the biggest challenge in building models over the course of your career?
A: The biggest challenge in creating predictive analytics solutions is ensuring that the target variable is created correctly. This is easier said than done as the practitioner has to do the following:
- Identify information to be used in creating target variable
- Creation of analytical file with pre-period where all information represents potential model variable inputs to model and post period where the only information is the actual target variable.
- Many problems arise when information in the pre-period represents a portion of the target variable or the pre-period overlaps with the post period. These kinds of seemingly simplistic data issues represent the lion’s share of problems when it comes to overstatement of model results.
Don’t miss Richard Boire’s presentation, Predicting Extreme Behavior to improve the rating structure for Travel Insurance, at Predictive Analytics World San Francisco, on April 1, 2015 from 4:40-5:00 pm. Click here to register for attendance.
Richard Boire will also be providing his new book, Data Mining for Managers: How to use Data (Big and Small) to Solve Business Challenges, to those who attend his session at PAW San Francisco. His book provides streamlined insights chock-full of engaging stories, case studies, and techniques for making the most of the masses of information and mining techniques that technology has enabled.
Enjoy his session and his book by attending PAW San Francisco.