What Makes Predictive Analytics Work

What makes predictive analytics work? Good data, great algorithms, smart statisticians? Yeah sure, that stuff helps but none of it makes predictive analytics work. In my experience the most critical thing in making analytics work is a an operational business leader with the experience and vision to see them integrated into their business. I know this sounds like the old adage about CRM solutions not being about the software but being about the people and processes. In fact it is, and maybe doubly so for predictive analytics and decision management because the whole goal of an analytics solution is the tell operations (marketing, sales, customer service, collections, etc...) the right decision and get them to do the right thing.

Last week I got a call from a client. This experienced operational manager, Scott C., understands what makes analytics work and how to apply them to make his operations work better. Instead of waiting for his analytics group to suggest uses for predictive analytics Scott keeps a constant eye out for decisions that he and his team make that could be improved (made more profitably, faster and more consistently) through predictive scores and decision rules.

Scott's called last week was about a situation with his current collections and recovery group. Like the settlement offer pricing  and outsourcing decisions we've helped him with in the past, Scott is looking for an application that will tell his managers which accounts are right for a specialized treatment they've developed. The catch of course is that this treatment, while effective, is expensive. Our goal is to develop predictive models and strategies that optimize this new treatment's use given the bank's goals and constraints.

In this way Scott is leveraging analytics, not to replace operations but to supercharge it. Scott knows that only his operational team could have come up with the new treatment but that only analytics can prescribe it's use for optimal impact. When analytics are embraced by operations in this way new solutions are quickly developed for decisions that the analytics side of the business may never have known existed. What's more, because these new solutions are directly aligned with the operation's goals (bonuses) they are often better understood and more quickly adopted.

Intelligent Enterprise Magazine: Performance Notification Is Not Performance Management

Intelligent Enterprise Magazine: Performance Notification Is Not Performance Management

This is an interesting topic for those interested in Triggers. It begins to talk about the complexity required to make triggers, or event-based operational intelligence useful for solving real business problems. I posted a bit more on this in Customer Analytics and Decision Management: Triggers in a data driven world.

Scores only work if you use them

If you've been in the analytics business for a while you've had to deal with customers asking about the business value they are going to get from a new score or from scoring in general. The quick answer should be a question, "How are you going to use the score"? Only once the strategy is understood can we value the score. In other words, don't just look to scores and models to add value to the business, instead look to an entire strategy.

The case study below is based on work Don Davey, Director of Collections and Recovery Solutions at Intelligent Results has done and illustrates what I'm talking about. For the first half of the example Don uses a very simple strategy where the client increases 30 day collections by about 8% through differentiating actions across segmented populations instead of treating every account the same. The blue dots below provide more details about this business case.

1. Historically, accounts like these have had a 30 day collections value of $1,145,008 and a 30 day liquidation rate of 0.9% when all worked together.



2. The client splits the accounts into two nearly even groups based on the IR 30 day payment score. This score rank orders accounts based on their likelihood to make a payment. Therefore, low value accounts get low scores and high value accounts get higher scores.



3. This separation between the segments is apparent when you look at the calculated 30 day collections ($188K vs. $1M), liquidation rates (0.2% vs. 1.7%) and unit yields ($1.63 vs.$14.64) for each segment.



4. The strategy was to work the higher value group 50% more and thereby increase production from that group. It turns out that adding 50% effort to that segment yields an increase of about 10%, moving 30 collections from the high value segment from $1.026M to $1.128M.



5. The opposite treatment was performed on the low value segment and the result is a 10% reduction in their output, lowering collections from $119K to $107K.

Ultimately the increase in production from the high value segment greatly outweighs the decrease from the low value segment and in total the new strategy yields an overall increase of 8%.

Hopefully, this simple example illustrates how it's not just the score that creates business value but in fact the score combined with a solid strategy of differentiating actions. In reality you would seldom split your portfolio down the middle because a bit more analysis can tell us where to draw the line and which accounts warrant what levels of calling effort. In the next posting we'll get into how to really juice your calling strategy by optimizing effort levels for multiple segments based on call and agent costs, capacity constraints, and the expected values of different effort levels for each account. Lastly, before you jump in and start slicing up your operations please make certain that you've got the right infrastructure to plan, simulate, test, execute, track and refine data-driven strategies.

Tags: modeling, scores, collections, recovery, decision engine, rules engine, predictive modeling