Image
Math Students Win Insurance Company’s Modeling Competition
Two doctoral students in mathematics and statistics, Tangxin Jin and Jie Wang, recently brought home the first place trophy to campus after winning the annual Travelers Analytics Case Competition held in late January at the University of Connecticut Storrs.
Sponsored by the insurance firm headquartered in Hartford, the competition presented 71 teams of students from UConn’s Storrs and Hartford campuses as well as UMass Amherst with a real-life data set of around 20,000 records with 25 variables each, including such factors as accident reports, driver information, claim details, vehicle information and location of accident. Their task was to use probability theory, predictive modeling and machine learning to predict the likelihood that any given auto accident claim was fraudulent. As Jin explains, “Fraudulent claims are a big problem for insurance companies and mathematical predictive models can identify cases for further scrutiny.”
After months of elimination rounds, three finalist teams, one from each campus, progressed to the last round and were judged by insurance company officers and members of its actuarial and advanced analytics teams on presentation of their data exploration, model accuracy, cost assessment and business value.
Jin says, “It was a great honor for us to bring the trophy back to Amherst. The competition is annual, and it is helpful for students in our department to apply our diverse training and at the same time gain modeling experience by tackling real-life problems. We learn and develop mathematical and computational methods in classes and during our Ph.D. training, but here we applied these tools and training in this competition. We both feel we learned a lot from the real-life data and real-life problems.” Each hopes to pursue a career in mathematical sciences in industry.
Wang explains that dozens of teams entered the first phase of the competition last October. She and Jin used advanced machine learning approaches such as Random Forest, Bagging, Adaboost and XGBoost to build variant models. Then they combined several models using ensemble learning methods such as stacking to get better prediction performance. Based on the resulting predictive models, they found a few important factors such as driver’s safety rating, the claim’s estimated payout and vehicle price had a large influence on the likelihood that a claim was fraudulent. The company’s data scientists said they gained new business insights from Jin and Wang’s new “models of models.”
Mathematics and statistics department head Nathaniel Whitaker says of the team’s accomplishment, “The department is very proud of Jie and Tangxin in winning this competition. This highlights their talent and training. They have a broad range of expertise which will serve them well in the future.”
Professors Markos Katsoulakis and Luc Rey-Bellet say that they are very proud of their Ph.D. students’ accomplishment, but, Katsoulakis adds, “I am not really that surprised. Jie and Tangxin have mustered a wide-ranging expertise combining applied mathematics, statistics and computing that allowed them to be at the same time creative and practical, ultimately bringing them in at the top of the competition. This showcases the training of our applied mathematics students, but specifically demonstrates the strong potential of Jie and Tangxin in any future career paths they may choose.”
Asked how difficult was the competition, both students said the experience was indeed difficult but also fun. The one stressful element, they note, is that they could see their team’s ranking against the other 71 teams each week, and it became increasingly hard to stay on top. In addition to the trophy, the first place team members each brought home certificates and a mini-drone.