We make predictions and act on them all the time. I predict that if I jump into the path of a moving bus, I will be hurt – so I won't jump. I'd conclude that my prediction had been in alignment with my goals, but if I had to, I could only prove it by using the laws of physics or examples of other people's encounters with moving buses.
If done well, predictive analytics help companies avoid business situations analogous to being struck by a bus. Business situations, however, are usually less dramatic and much more nuanced than avoiding a moving vehicle. And, unlike the bus, a company will often not even know there was a situation worth avoiding.
Even so, business peril requires us to try to stay ahead of trouble. Predictive analytics are key to the prevention of loss by fraud, churn and other bad outcomes. Predictive analytics also help prevent the loss of wasted time and money spent on activities that do not contribute to business goals.
But there are limits to the usefulness of predictive analytics as we have applied them to date. One conclusion we have reached is that it is no longer sufficient to simply try to predict an unimpeded future. We must hedge our predictions with probabilities and be aware that a variety of reactions to those probabilities might be in order.
Many predictive models are tuned to report a binomial result, for example, "likely to churn." In practice, multiple actions could occur as a result of this discovery, including "do nothing." Whatever the reaction is (even to an event that has not yet taken place), it must be in alignment with company goals. The predictive models are important unto themselves, but I will focus here on how to support the actions we take when using predictive models, the "next steps" that are often neglected.
Predictive analytics are applied in the process of determining business events that are likely to occur and actionable. The probability threshold of "likely" differs from company to company and risk factor to risk factor. A risk-averse company may decide to prepare for a relatively low probability event that comes with a particularly bad outcome. Companies that are more risk tolerant, unaware or distracted by other projects will be less poised to take action. (See Figure 1.)
The reality is that most companies that do predictive analytics do so without attaching a probability. A predictive modeler might produce a result that indicates a customer is likely to churn, yet the model might not indicate how likely or whether the company should care if the customer leaves. Probabilities and commensurate tiers of reactions need to be in place if we wish to fully utilize predictive analytics. Increasingly, predictive analytics users are taking this more-nuanced approach, and analysts are becoming attuned to their environment and their master data, which should also influence the action taken.Let's examine some of the real decisions that need to be made in the many things predictive analytics can refer to:
Example 1: Customer Lifetime Value
Customer lifetime value is a means to an end. It supports operations as a data point to justify taking other actions, such as whether to market to a person/company, how to support the customer, whether to approve a loan, whether to challenge a transaction as fraudulent, etc. I'm including CLV here to be consistent with prevalent practices of predictive analytics and to note that it should be forward-looking, which is not profit-to-date, but projected profit over the next few years.
Example 2: Clinical Treatment
Care-giving organizations want to provide the best care at the lowest cost. To reach this balance, in principle, multiple procedures for the patient are considered based on probabilities of efficacy. Procedures have subtleties, which shows the need for predictive analytics that start with the customer, not a campaign.
Example 3: Churn Management
When a customer appears likely to churn, companies are increasingly turning to customer lifetime value and other predictive analytics to temper the instinct to rush to the account representative for an attempt to salvage the relationship. The operative term is "churn management," not "churn prevention." (See Figure 2.)
And, sorry to say, people like me who just pay their bills every month do not necessarily fall into the highest CLV and highest salvageable category. (This is painfully apparent when I experience exhausting hold times when trying to reach customer service.)
Regardless, proactive intervention to salvage the relationship, if so desired, is a multidimensional decision. Figure 3 shows one of the factors that probability to churn needs to be juxtaposed with: customer lifetime value.
Example 4: Segmenting for Next Best Offer
Descriptive modeling classifies customers into segments that are utilized in a large variety of marketing-related activities. These segments should be formed dynamically in conjunction with campaigns and should correlate to the various activities of the campaign. Rather than marketing to everyone determined "likely to purchase," a "probability to purchase" should be produced and used with other factors that make the effort worthwhile to the company in the long run. A factor like the customer's income or another decile might increase the company's interest in encouraging the customer through a campaign offer.
A related use of predictive analytics is in decision modeling, which might focus on the next customer interaction and whether it should be proactive and driven by the company (like extending an offer) or reactive (like responding to a loan application).
Example 5: Fraud Detection
Predictive analytics is used to determine the potential fraudulent nature of a transaction. Here again, we find that analyzing a transaction without considering how likely it is to be fraudulent and without bringing to bear a customer profile of summarized and recent transactions can lead to false assumptions and actions. Increasingly, a customer profile is required reading for any decision engine performing fraud detection. We find suitably robust profiles primarily in environments that have adopted master data management.
The human element does not disappear in the use of predictive analytics. The trend is to maintain a level of human judgment through self-service predictive analytics, allowing the analyst to consider multiple variables and actions.
But in a larger and broader scale, another trend is bringing more data into the predictions, and that includes Web-scale data in Hadoop and other big-data environments. For example, blogs, now finally being captured in Hadoop, can contribute to CLV by quantifying customer usage burdens on support and other activities that produce probabilities.
Proving the need for multiple dimensions in predictive analytics is like proving I should not have stepped in front of that bus. It's sometimes hard to demonstrate what you have prevented. These techniques do not necessarily reduce churn, improve results from a given procedure or increase fraud interventions. You will find, however, that they do improve the bottom line, which is the higher calling for all involved.