Murli Buluswar, Chief Science Officer for AIG-Property & Casualty gave a talk in the Big Data Lecture Series -- Fall 2012 at MIT CSAIL. In this talk he shared details on how the newly created Science Office in AIG is starting to shape empirical based decision making in a broad suite of areas including underwriting, claims, marketing, sales, and even human resources. His entire talk revolved around the theme of applying data science to influence business decisions.
AIG property and casualty division is a $33-35bn business, spread across 92 countries. Buluswar explained insurance as selling a promise/service/product without knowing what the cost would be, and how data science can be useful to influence the cost and justify the investment. He mentioned three important goals of the data science team in AIG -- forecasting (what can happen), predictive modeling (the underlying phenomena that explains why something would happen) and optimization (the best outcome).
He emphasized that while the insurance industry is data rich, the ability to mine the data by asking questions is antiquated and thus it tends to be knowledge poor. Giving an example of auto-insurance, he claimed that the factors that insurance companies currently take into consideration -- age, gender, marital status, accident history, income, profession, educational background, etc. are not the core characteristics which can explain the risk; the characteristics like how does one drive, where does he drive, when does he drive, how much does he drive, what vehicle does he drive etc. explain the risk better and a pricing model based on these behaviours would be far more effective.
In U.S., the insurance pricing need to be justified by explaining all the factors (for an example, previous driving experience) and how these factors influence the risk and thus eventually justify the pricing. The insurance companies can’t create a black box model; they need to show in a transparent way, the connection between the characteristics and the pricing. This justification is possible only by mining data.
Data science can be very useful across all the business functions - pricing, marketing, distribution and claims (fraud detection). Taking auto Insurance as an example, Buluswar explained that different groups respond differently to price change and using historical data, they can predict how would customers respond to a change and how would their competitors behave. To detect fraud, they use social data like why would somebody go to a medical service provider 100 miles away. In summary, they use data to augment human judgment. They use weather pattern, catastrophe modeling, satellite images etc. to predict natural disaster.
They use data science to manage the intermediaries (brokers) -- quality of customers they attract, the margin they take etc. They use data science to influence the spending -- how much on advertising, how much on promotions, online, against competitors; how different media (print/online/tv) give different returns and how return on investment can be optimized. By mining various kinds of data, their data science team influence every aspect of the business, in a positive way.
In the end, he gave a very interesting example that demonstrates how useful data science can be in understanding some subtle characteristic patterns like role of fatigue in accident, the correlation between road vibrations and fatigue, and how a truck equipped with a good vibration cancellation system is less risky than one that doesn’t have a vibration cancellation mechanism.