Forecasting Customer Purchase Patterns via Enhanced Data Metrics

Forecasting Customer Purchase Patterns via Enhanced Data Metrics


In a business setting, the purchase history of each customer is typically summarized via three metrics that include the elapsed time since their most recent visit (Recency), how many purchase visits they have made overall (Frequency), and the total dollar value they have spent to date (Monetary). These values are referred to as RFM, and are computed for each individual when forming segmentation groups in a customer database, as well as during estimation of Customer Lifetime Value.  However, many researchers have expressed concerns that RFM metrics alone are insufficient, and have suggested that the framework should be extended to include several additional measurements. In particular, standard RFM strategies do nothing to reveal important patterns of behavior in the time-series history of the customer. Unfortunately, the various additional metrics are scattered in the research literature, and no collective assimilation seems to exist, especially in the form of software tools.



To increase analytical precision, AlgoTactica has designed object-oriented database software and enhanced data structures that track metrics for every time point in a customer’s purchase record. In combination with RFM history, our enhanced data objects also encapsulate time series patterns for 16 additional metrics pertaining to customer behavior. These metrics can be queried to form extended feature sets that enhance predictive model capability and offer new segmentation strategies for Customer-Base Analysis. One very important new metric that we use has been proposed by marketing researchers in order to identify customer repeat visit patterns that depart from regularity, and which thus exhibit intermittent or clumpy behavior across the purchase time-series history. Here, we discuss some aspects regarding the use of this metric, based on analysis of 2000 customer histories in a grocery retail setting.



For the customer cohort considered in this study, a full 25% of the group was identified as exhibiting clumpy visitation behavior. Consequently, allocation of these customers to a separate segmentation group can offer significantly improved insights for sales anticipation and planning of direct marketing activities. In particular, the Days Between Visits graph shows that regular-pattern customers will typically engage a repeat purchase once every 5 days, and overall most vary within the range of 1-15 days. However, clumpy customers do not seem to exhibit a dominant typical value, but on average will visit once every 13 days, and most of the overall group will vary within the range of 3-23 days before engaging a repeat purchase. Customers within this group might very well be sharing their purchasing with other businesses, and once identified can then be engaged via direct marketing efforts that entice more regular visitation. Furthermore, the Spend Value Per Visit graph reveals the potential revenue loss due to clumpy behavior. At the 90’th percentile, this graph shows that regular customers spend upwards to $65 per visit, whereas clumpy customers typically do not exceed $50 dollars per visit.

If undetected, clumpy behavior can also lead to misinterpretation of certain other vital metrics. For instance, marketing actions aimed at loyalty retention are often driven by measures that indicate a low probability of the customer still being active with the business. In this study, it was known that all customers, including the clumpy ones, were still active purchasers. However, as shown in the Probability Customer is Active graph, the clumpy customers were often scoring much lower time-averaged measures on this metric than the regular customers were, leading to a possible misinterpretation that they might have churned.  For instance, with reference to the x-axis it is seen for clumpy that approximately 50% of the area under the curve is to the left of the 0.8 probability of being active, whereas the curve for regular has only approximately 20% of its area to the left of this point; yet, over the duration of this study they were all repeating purchasers. Therefore, when evaluating propensity-to-churn metrics, clumpy prior behavior of otherwise loyal customers must be taken into consideration. In cases of lower scores, a customer might merely have entered one of the temporarily-dormant phases that are typical of clumpy behavior.

Back to top