How to Forecast What New Customers Will Do Next

April 3, 2017

Brand-new customers pose an interesting challenge to marketers in this era of big data. Marketing strategy and tactics are often driven by valuable insights gleaned from past customer data collected over time from repeat purchases and transactions. But newly acquired customers arrive without a data trail. Is it possible to predict the future behavior of new customers?

As a doctoral student in marketing, Arun Gopalakrishnan seized on this challenge to create a new computer model to provide guidance to managers who want, and need, to forecast newly acquired customers’ behaviors. Gopalakrishnan began this research as part of his doctoral studies at the Wharton School of the University of Pennsylvania and completed the research at Olin Business School.

“Our findings suggest that simply using older cohorts [sets of customers acquired in the past] as a proxy for predicting new cohorts without understanding any potential regime changes may lead to inaccurate predictions because certain aspects of customer behavior may have changed, going from the oldest customer cohort to the newest one.”

Working with two professors at Wharton, Gopalakrishnan developed two cross‐cohort models (called vector changepoint models) that introduce a new framework for analyzing data that reveals insights into patterns of customer behavior over time. Specifically, the new models reject the notion that pooling data from all previous customers to make an educated guess about the behavioral patterns of the newest customers provides an accurate forecast. In other words, the researchers found that new customers are not simply going to behave like the “average” existing customer. That assumption, according to the researchers, “ignores the potentially changing behavioral patterns” from one set of customers acquired during a certain time period to another.


The new mathematical model takes into account what it calls “regime changes” or past customer behavior changes that were influenced by new firm policy, government regulations, economic factors, competitors’ actions, or unknown drivers of change.

KEY TAKEAWAYS for Managers

  1. Simply using older customers as a proxy for predicting new customer behavior without understanding any potential “regime changes” (influences that shift or affect customer behavior), may lead to inaccurate predictions.
  2. The vector changepoint (VC) models proposed in this paper eliminate the need to run thousands of possible models to determine the “best fitting one” which is often a problem for marketers.
  3. Data analysts will be able to use the VC models to yield robust predictions of cohorts’ future behaviors and tease apart calendar time effects as well.

The model was applied to a data set from a public television broadcaster in the United States that relies on voluntary donations for part of its operating budget. Comprehensive data from more than 50,000 donors who were acquired over a 10-year period were used to estimate the model parameters. That time period included the passing of the Telecommunications Act of 1996, which the model suggests is a type of regime change that may be driving changes in the behavior of donors acquired in or after 1996.

“Our insight is you have to use the right data for forecasting what the new customers are going to do,” explains Gopalakrishnan.

“That doesn’t mean there is bad data, but we have found that what a customer does is more important than demographic information. And in our data setting, the evidence suggests that new customers are going to behave differently from ‘old’ customers’ behavioral patterns.”

When tested against other models, the Olin Award–winning research model/forecast tool outperforms other benchmarks. It can be applied to any industry that acquires customers who engage in repeat transactions over time. The new model also simplifies the process of mining the data.

According to Gopalakrishnan, “There is no need to run multiple models to figure out what fits best—our model allows for a large number of possible regime changes, including none, and can determine which model best fits the data in one unified framework.”

Prof. Arun Gopalakrishnan presented his research to alumni and friends in the business community on April 27, 2017.

“A Cross-Cohort Changepoint Model for Customer-Base Analysis”


Arun Gopalakrishnan, Assistant Professor of Marketing, Olin Business School, Washington University in St. Louis

Eric T. Bradlow, K.P. Chao Professor of Marketing, Statistics, and Education; and Peter S. Fader, Frances and Pei-Yuan Chia Professor of Marketing, at the Wharton School, University of Pennsylvania


Marketing Science, Volume 36, Issue 2, March-April 2017