Making Data Shovel Ready for Predictive Analytics Projects 

Many times there is a gap between conventional wisdom and what actually goes on, he points out. "For example, an organization might be losing customers faster as it increases prices. The gut reaction is to blame the price increase, but analytics could show even the customers who did not get a price increase have high attrition levels similar to those who did.

"The real reason for increased attrition, therefore, would be something else. That’s the first ‘aha’ analytics is able to provide—differentiating between what is perceived vs. what is really happening in the business," says Arora.

The second level of understanding, continues Arora, is the ability to differentiate between causal and co-relational impact.

"Often, decision-makers confuse the two," he says. "For example, customers who pay in full have higher retention compared with those on monthly payments. However, will a customer’s retention improve if you encourage a monthly payer to move to paid-in-full status? The question is, does paying in full cause retention to improve, or do people who tend to stay longer also tend to pay in full? The latter implies that this situation is co-relational—they exist together because of certain other characteristics of the customers. When one is able to separate the impact of one thing over the other using analytics, what is left is sometimes very counter-intuitive and that’s the ‘aha’ moment."

Smallwood indicates one of the primary areas where predictive analytics would benefit the insurance industry is in the area of market intelligence—aligning products and distribution channels with the right customers.

That encompasses in underwriting everything from risk appetite to risk analysis and pricing precision, and on the claims side, at a minimum, fraud detection. "What it does for underwriting and claims can allow for more automation," says Smallwood. "Not only can it help reduce expenses, but what it ultimately does is improve the overall quality of your book of business."

"When analytical teams use a lot of detailed statistical measures and graphs in their communications, it can distance the business group even further," he says. "A good way to get the business on board is to predict the results for the next few months, share them now with the business users, and then validate them once they materialize. That litmus test is far more powerful than any statistical proof. Of course, you cannot use this for long-term predictions or for things that take many months to realize."

The second key is to concentrate on goals. "You should be able to narrow your modeling world into what you truly need to target," says Arora. "It is much more difficult to have a model that predicts accurately throughout the population than a model that is more tailored to your objective. The analytics team needs to understand what will drive the decision and concentrate on improving the model at that tipping point."

The third key is to understand the data. "You need to be careful about things such as what source to use, what time period, and what kind of sample," says Arora. "For example, sometimes there are one-time events that occurred in the time period you selected and have a big impact on the outcome. If you fail to identify and normalize for these, the model predictions will be skewed.

"It’s one thing to make a good model and another to make a good model that can be easily implemented," he continues. "A lot of times people make good models, but once they get to the implementation stage, they have to compromise, and [the models] lose a lot of power." Arora maintains the 80/20 rule applies to analytics, as well. "You could need 100 variables in a predictive model to get great results," he says. "You also could get 80 percent of the results by using just 10 variables. Though it finally depends on what you are trying to go after, more often the simpler model is the better choice. It is going to be a lot more stable, a lot less prone to variances, a lot easier to implement and validate, and it still is going to drive a whole lot of value."

Many companies actually hire people from outside the insurance industry, Smallwood points out, and bring them into the department to plant the seed and grow that expertise in-house.

"What I’ve seen is insurers are going to the retail market, where marketing campaigns with predictive analytics have been done, and they bring someone from there into their own marketing department to help with that distribution channel," she says.

"The retail industry has been doing [analytics] with catalogs, online, stores, and advertisements. There are not a lot of people who know [predictive analytics] in insurance unfortunately," she says.

Arora is an example of an analytics professional whom Smallwood described. He began his work in analytics with Capital One before joining Farmers. "Insurance seems to be lagging behind in use of advanced analytics. If you think about it, insurers were the ones that started using the power of data earlier than anybody else. But other industries started at the next level while the insurance industry stayed with its 100-year-old methodologies," he says.

Smallwood adds you can buy that knowledge, but you still somehow have to transition it and have a training program.

However, Carney predicts, as carriers hone their skills and their ability to apply the data to their business, analytics is going to become even more valuable in terms of the insights it generates.

"I absolutely could see an analytics czar within the organization," says Carney. "It could be an interesting role. As people get more comfortable with the interpretation and application of [analytics], you could see specialties emerge that could be a wonderful career track for someone within the organization. I would suspect in terms of the paucity of skills around fraud, this could involve a high skill set for people."

When asked about which is easier to develop—good data or a sharp staff of analytics professionals—Arora replies he doesn’t believe one can come to an easy conclusion.

"They are both challenging in their own right," he says. "Neither is an easy task. Most often you will see both need to be developed simultaneously. You need great analytical people to help acquire and maintain good data, and you need good data to engage the analytic team to do what it does best."

But the long wait makes it imperative insurers get it right from the beginning. "For us, knowing it takes longer for things to develop and see results, we can’t afford to miss and immediately turn around and refile," says Isakson.

"We spend a lot of time on the validation upfront before we deploy the model through our book of business and [rely on] historical results to make sure, for what we are deploying, we can predict how it will behave," he adds.

At Pinnacol, the models on the underwriting side are studied annually. "Unless your market is dynamic and changing dramatically, you probably are building something you want to monitor; but give it some time first to see the impact before you start tweaking it," recommends Isakson.

"You really need to allow it to flow through your book of business, so a window of two to three years probably would be ideal before making substantive changes unless you can’t avoid making changes," he continues. "You may have to make refinements here and there along the way but probably not much. We haven’t had to make any refinements yet because [the model] has behaved like we expected during the evaluation. On the results side, we are starting to look at that because we are two years in and should start to quantify the impact from an outcome standpoint. If you are constantly tweaking it, you don’t have a predictive model."

Tweaking the models is not a constant process, but Carney believes it is a regular occurrence. "As your business changes and you identify new behavior or characteristics or the tool reveals something you don’t expect, you have to react," she says.

A lot of insurers have addressed data issues through what they’ve done with policy administration systems over the last few years, remarks Carney. Now, they can fix other areas because they are more confident in the data quality throughout the enterprise. However, "if they’ve not gone through that exercise, the data has issues, and they are not going to be very confident of the outcome," she warns.

Size matters, too, Carney continues. "Smaller carriers are definitely looking at [analytics] and following the tier-one carriers," she says. "When you are competing with a State Farm or an Allstate, you have to use technology to help you overcome their size and skill. A lot [of smaller carriers] are looking at predictive analytics, and if they can have it delivered like SaaS, that can overcome some of the cost hurdles for smaller carriers or those that are focused by region or line."

The issue always seems to come back to the data, though, and Smallwood maintains until the business side really understands insurance is data-centric, it’s going to remain an uphill challenge.