A: We see a wide variety of models, predicting a wide range of things, but all the predictive analytic models we work with are being used as part of a high-volume, automated (or partially automated) decision. The model might be predicting a customer’s propensity to accept an offer and be used in a decision to pick which offer to display to them (next best offer). Or it might be predicting the likelihood that a customer has not disclosed a pre-existing medical condition as part of deciding whether to auto-pay a claim or not. The predictive analytic models are useful, sometimes essential, elements of a broader operational decision.

Q: How does predictive analytics deliver value in your client’s organizations – what is one specific way in which it actively drives decisions or operations?

A: One client used a predictive model to decide if a particular medical report did, in fact, state that the person had the specific medical condition they were claiming disability for. This was used as part of a decision to fast-track a claim or not. Fast track claims have a lighter weight and cheaper process (unless they find something odd). The predictive model replaced a manual review of the document in those claims where the policies, procedures and regulations implemented in business rules allowed for fast tracking as long as the medical report agreed. This let them increase the number of fast track claims, without fast tracking the wrong claims, and made the process more effective and more customer-centric (customers strongly prefer the fast track process).

Q: What’s the difference between the predictive lift of a model and the ROI of an analytics initiative?

A: The predictive lift really tells you how good the prediction is – how much better than average the model is if you like. But predictive models are not the whole solution when it comes to creating business value. For instance, a predictive model with good lift in marketing is giving you an accurate assessment of how appealing a particular offer is to a particular customer. That’s great. But your business objective is not to predict propensity to accept offers but to get customers to accept offers. And there may be regulatory, policy or contractual reasons that mean you can’t just offer the most appealing offer. The customer may own a product that means they can’t buy certain other products or may be ineligible for the product or there may not be capacity right now to take on more customers for the product. The decision as to which offer to make must consider all these things. And the ROI is going to come from the effectiveness of the decision as a whole, not just from the lift. The lift may have an impact but its not the only element that matters.

Q: What surprising discovery or insight have you unearthed working with data?

A: That data scientists spend too much time with their data! Seriously, they love working with data so much that they would (mostly) rather do that than talk to business experts and IT professionals. Yet the business experts are the only people who can tell them what kind of prediction will be useful (and used) and IT knows where the data is, what the challenges are likely to be with it and how the result will need to be integrated into operational systems. A little less time listening to the data and a little more time listening to the business would improve results!

Q: Sneak preview: Please tell us a take-away that you will provide during your talk at Predictive Analytics World.

A: In most companies, the value of predictive analytics comes not from influencing the decisions made by executives or managers, but by embedding predictive analytics in automated, digital decisioning systems that make decisions about single customers, single interactions, single transactions.