Discussion with State Farm's Eric Webster: Insurance and Data Mining

Although I missed SAS Global Forum, SAS kindly arranged for me to talkto some of their clients about analytics. Here I present my discussionwith Eric Webster, vice president of Marketing at State Farm.

Eric WebsterEric Webster joined State Farm in July 2000 as assistant vicepresident, Customer insights, Marketing. He was promoted to his currentposition in April 2007. Before coming to State Farm, he was vicepresident of FCB Direct in Chicago, and previously was senior databasemarketing manager for IBM, responsible for marketing analytics for IBMworldwide. Eric has BS in Finance and an MS in Math from the U. ofIllinois, Urbana-Champaign.

Gregory Piatetsky-Shapiro: What are 2-3 most important things you do with analytics / data mining?

Eric Webster:One of the major applications is direct marketing campaigns.

Our analytics enable us to understand customer purchase cycles -how likely are they to leave us, what the customer is likely to buy,what is the next product to offer.

We also give tools to our agents to enable them to use analytics results without having to conduct analyses themselves.

Another major area where we focus is actuarial - where data mining is central.

Making sure that we are matching the right price with the rightrisk is the core of what insurance is, and that is an analyticexercise.

Understanding who is most likely to get into an accident, andgenerating the right price for them, and at the same time helpingpeople who are more at risk to lower their risk.

For example, we can look at teen drivers, who often haveproblematic driving records, and provide discounts for them if theytake safety courses.

Another major area is what we call risk management. Our portfolio of risks at State Farm includes thinking about risk:

• what happens when the next hurricane comes along
• what will that mean in terms of potential financial loss
• how many claims reps we will need to send to any particulararea, given, for example, that a force 3 hurricane will make landfallin this particular spot.

So even before hurricane makes landfall State Farm claimforces are already on their way, and they know how to dispersethemselves in the region.

GPS: When I worked with insurance data, one of the problems was thatinteractions with customers were very infrequent - sometimes once ayear. It was difficult to build good models using such slowly changingdata. How do you deal with this problem:

Eric Webster:It is an issue. But insurance itself is a slowly changing dimension [so slow changes in data are not a problem if the change in risk is also slow. GPS].

Most people experience changes in their needs for insurance based onlife events such as getting married, having children, buying a car, andmoving, which do not happen frequently.

GPS: Where do you currently see the highest ROI from analytics?

Eric Webster: Any advances we make in matching price and risk is huge.

If we can lower the prices for certain groups of people withgood driving records, that increases their customer satisfaction andincreases our ability to keep customers.

From a pure marketing perspective we use analytics to predictwhat is going to happen in our marketing campaigns. We let theanalytics drive the campaign - we don't mail to people who are notlikely to respond, etc. I cannot quote numbers, but we have seriouslyincreased the ROI over what we were doing before we were doing datamining.

GPS: Where and how analytics need improvement to meet your current needs?

Eric Webster: In terms of improvements we would like inanalytics tools is to make them easier to use - have the tools makeintelligent defaults for us.

For example, if I use a neural net node in SAS EnterpriseMiner, I would like it to choose its parameters and algorithmsadaptively - have it train itself on the data for a given problem.

GPS: More automation?

Eric Webster: Yes, and have the tools help us make the rightchoices. So I don't have to always have PhD statisticians running thetools. Which convergence criteria should I use? Should I use step-wiseregression variable selection or backwards/forwards selection? GLM orlogistic regression?

Can we tell the tool - here is a subset of the data - go play with it and tell me which methods and algorithms should I use.

[GPS: there are tools which have higher degree of automation, for example KXEN;
SAS also offered to discuss their plans for increasing automation, which may be a good topic for discussion]

GPS: Do you use text analytics?

Eric Webster: we experiment with it, but it is not currently in production.

We get a lot of comments written on our website, so we look attools that can do "triage", to identify which ones which humans shouldlook at. This one should go to a claim rep, this should go to themarketing dept, this should go to the website people. It is not inproduction now, but it is something we are looking at.

GPS: What role can analytics play in the current financial crisis?

Eric Webster: From our perspective it is important to getfundamentals right. Make sure that we keep a sharp eye on our costs,making sure that we are as process efficient as we can be, and the waywe do that is largely thru analytics - where are the areas where we canimprove, can we tweak the models and increase the ROI. The currentcrisis demonstrates the need that we stay on top our game, andanalytics is at the core of that.

GPS: You may have a Wall Street Journal article (Jan 26, 2009) that listed the top 200 jobs.
#1 job was a mathematician, #2 was actuary, and #3 was a statistician.
Actuary has a reputation of being a boring job. Can you tell us what actuaries actually do and why their job is so great?

Eric Webster: My degree is in math, so I can relate to all ofthese. Some people think, "Can there be anything more boring thaninsurance?!?" But when you start to dig into what it is all about - yousee that there is a lot of interesting work to be done, there is anever ending demand for analytics, and insurance is one of very fewplaces where data mining and analytics turn immediately into companyfundamentals. This is what actuaries and statisticians do at StateFarm. It is a very exciting profession for people who like data miningand numbers, since people really care about what you produce.

Insurance is nothing but management of information. It ispooling of risk, and whomever can manipulate information the best has asignificant competitive advantage, so I agree that a mathematician,actuary, or a statistician is a safe and rewarding job.