The use of predictive analytics in marketing is critical for understanding and predicting the "time" dimension...the holy grail of marketing is to be able to make the right offer, through the right channel, to the right customer/prospect, at the right time. The most critical, and arguably the hardest part to get right is the time dimension. An understanding of time can help answer key questions such as: When is my customer/prospect most likely to purchase or be in-market? or How long before my customer ends their relationship?
In addition to predicting time itself, it is also important to understand its drivers. By doing so you can find answers to: What can I do to shorten my customer's buying cycle? Are there certain experiences that actually accelerate time to attrition?
Answering these questions requires a deeper analytic knowledge and application of specialized econometric techniques such as the development of Hazard Models - also referred to as survival models, duration models, or failure time models.
Traditional Analytic Approaches Fall Short
Since time is a continuous variable, and regression methods are well developed, why do we need specialized techniques for prediction? Traditional approaches fall short because people buy and sell at various times, and not all of those times may be observed in the data. In the example below, we are interested in predicting when an automobile will be replaced given that it's been owned for some time and has not yet been replaced.
- We have data on customers from 2005-2012 showing what they own and when vehicles were purchased/disposed in the past seven years. For some customers we may see that they have a vehicle that was purchased before 2005 and disposed sometime before 2012. Other customers may have purchased sometime during 2005-2012 but have not yet disposed of it. In general, longer ownership cycles are under-represented in this data, and traditional methods are not able to account for this bias. In order to predict disposal times, it is important to account for this bias. Not accounting for it would result in an under-estimation of disposal times, thereby leading to a missed opportunity for marketers to target customers when they are likely to dispose of their current vehicles.
- Hazard models are also well suited to tackle covariates that change over time. For example, a consumer's disposable income that changes over time may be highly predictive of their vehicle disposal timing. So, what value of disposable income should be used in the analysis? Disposable income in 2005? In 2012? Some average over the 7 years? This value will help us determine what changes in disposable income are likely triggers for the disposal of a vehicle, thereby enabling marketers to target consumers at the right time.
Why Are They Called Hazard Models?
In the above example, what we are interested at any given point in time is the likelihood that an event (vehicle disposal) will occur given that the event has not occurred thus far. In other words, we are attempting to determine the risk or hazard of vehicle disposal. Hence, the term "hazard" models. The econometric details of these models are too involved to go through in this post so if you would like more information about how to use hazard modeling to understand time to decision and its drivers for your customers and prospects, feel free to contact me at 781-494-9989 x201 or firstname.lastname@example.org.
About the author: Niren Sirohi is vice president of predictive analytics at Peppers & Rogers Group. Contact him at email@example.com.