What organizations want to know is not what kind of technology to buy first or what techniques and training they need, but what kind of problem to go looking for. What kind of problem will show the greatest return on an investment in predictive analytics? Where can they apply predictive analytics and get a clear and compelling "win"?
In my experience, the best place to start with predictive analytics is in your day-to-day operations. Operational decisions are about a single customer or transaction. Examples include: "What offer should I make to this customer to retain them?" "What loan can I offer this person?" and "Is this claim fraudulent?" Operational decisions are the best place to start with predictive analytics because they are transactional, because other approaches to information-based decision-making don't work well and because operational decisions align best with the potential of predictive analytic models.
As already noted, operational decisions relate to a single transaction. Every order placed, every loan application and every claim requires a decision. Because of their nature, it easier to build effective predictive analytic models based on transactions. The sheer volume of transactions means you have lots of data to build predictive analytic models with (and analysts like more data to work with). You'll often have data for a specific customer that stretches across time also, with each new transaction adding more data. Predictive analytic models often rely on analyzing the behavior of customers over time.
Because the decisions are transactional, only a finite set of defined actions can be taken as a result. For instance, you can approve a claim, reject it or refer it for investigation. This well-defined set of options makes it easier to tie results to decisions and thus create the clear feedback loop that helps you improve predictions over time.
Finally, the volume of transactions also allows you to experiment. You can try different cross-sell approaches with different customer orders to see what works best. This data also improves the quality of predictions.
Information Presentation Does Not Work
Operational decisions are made by nonexperts. They are made by front-line staff in call centers and retail outlets or by automated systems and websites. In the absence of skilled knowledge workers to review information, techniques such as OLAP or visualization are less valuable than they are for other decisions. Operational decisions must be made quickly, often in the context of a running transaction. Embedding a predictive analytics model that can be executed to "score" a cross-sell offer or loan product (based on how likely it is predicted to work) is much faster than presenting this same information to someone visually and asking them to interpret it. When no one is making the decision (as when a website makes a cross-sell offer, for instance) visualization is meaningless because no one is looking at the data anyway. Without the time or skills to interpret information, operational decision-makers are much more likely to effectively use information that drives a predictive analytic model than to consume that information any other way.
Operational Decisions Align with Predictive Analytics
The final characteristic of operational decisions that makes them ideal for predictive analytics is their alignment with the power of predictive analytic models. You can predict three things with predictive analytic models: You can use prior behavior to predict how risky a customer is likely to be, you can use past interactions to predict opportunities to grow the relationship with a customer, and you can use past fraud to predict how likely it is that a pending transaction is fraudulent.
The capability for predictive analytic models to predict risk, opportunity and fraud aligns perfectly with operational decisions. Most kinds of risk (like credit risk and delivery risk) are not accumulated in big blocks but one bad loan, one bad order at a time. Operational decisions make the difference in how an organization effectively manages risk.
Similarly, opportunity comes one customer at a time. Operational decisions that maximize the value of an interaction with a customer are critical. Even when organized crime rings are involved, fraud drains money from an organization, one credit card transaction, one fake medical procedure or one staged accident at a time. Only an improvement in ability to make a fraud detection decision at the operational front line can prevent fraud from getting a grip.
Focusing predictive analytics on operational decisions can seem counterintuitive. After all, predictive analytics can require a significant investment and those making the investment may be inclined to apply predictive analytics to strategic decisions.
The reality is that putting predictive analytics to work in operational systems offers a much greater ROI than embedding them in executive dashboards. Operational decisions generate the large volumes of data you need, and they let you improve this data by experimenting. They cannot be improved with traditional BI tools, and they align perfectly with the power of predictive analytic models. Get started with predictive analytics by improving your operational decisions.
James Taylor is the CEO of Decision Management Solutions and is the leading expert in how to use business rules and analytic technology to build decision management systems. He is passionate about using decision management systems to help companies improve decision-making and develop an agile, analytic and adaptive business. He provides strategic consulting to companies of all sizes, working with clients in all sectors to adopt decision-making technology. Taylor is a faculty member of the International Institute for Analytics and is the author of Decision Management Systems: A practical guide to using business rules and predictive analytics (IBM Press, 2011). He previously wrote Smart (Enough) Systems: How to Deliver Competitive Advantage by Automating Hidden Decisions (Prentice Hall) with Neil Raden, and has contributed chapters on Decision Management to multiple books. He is a frequent contributor to Information Management and writes a regular blog at JT on EDM. You can follow him at @jamet123