June 10, 2019

Finding the hidden gold in your dormant customers

Many retailers keep data on far more dormant than active customers.

The problem though often is that without the means to sift the potentially valuable ones from the rest, they tend to remain unreachable.

This white paper showcases work we have recently undertaken for a fashion retailer where we managed to build a model that effectively identified groups of customers that could be reactivated, and provide a very healthy ROI.

Developing a successful reactivation model for dormant customers in the home-shopping sector

This case study is important for all B to C companies selling directly to the public with large numbers of identifiable but dormant customers because it:

  • Provides an approach for identifying dormant customers who will be responsive to re-activation
  • Shows examples of the levels of response that can be achieved from customers hitherto classified as dormant

Background and objectives

Dormant customers were defined as those who had not ordered in the previous two years.

Within that ‘dormant’ group, and going back a total of seven years, we selected for the development of the targeting model those that had made at least one transaction in the earlier period and who could also be contacted by mail.

This pool of potentially reactivatable dormants represented nearly as great a volume as all the actives who had transacted in the previous two years.

The objective then was to use regression analysis to develop a model that predicted expected spend in a forthcoming season at an individual customer level.

Development process

Key to building any successful marketing propensity model is the data that is assembled for it.

In this case we used for our target variable the total spend of a ‘dormant’ customer in a recent historic season, with the predictor variables relating to any spend they had made in periods prior to that.

These target dormants had not been receiving catalogues so could be described as self-motivating. They represented some 7% of all dormants.

The predictor variables that were included in the final model included:

  • The product categories in which the person had purchased
  • Combinations of variables around recency, frequency and monetary value
  • The number and value of any returned items
  • Variables available from Abacus, an organisation that combines data from multiple retail brands and hence provides information on individuals’ actual spend in a sector

The resultant model deciles came out as follows:

This can be viewed as being a very powerful predictive model.

Subsequent model performance

Our client then used the model to compare higher ranking ‘dormants’ with lower ranking ‘actives’ and was able, after testing, to switch a considerable amount of their catalogue spend from ‘actives’ to ‘dormants’.

An example of the benefit from this exchange was the comparative response rates to catalogues; they were able to substitute ‘dormants’ for ‘actives’ at an indexed response rate improvement of 41%.

In addition, the overall dormant demand in season model was accurate to 0.6% across all dormants when comparing actuals to forecast.

Key learnings

For many retailers there are considerable numbers of ‘dormant’ customers that in fact contain a great deal of latent value.

The problem is to decide on which ones to try to reactivate.

However, introducing propensity models using this approach, and integrating industry wide demand variables from Abacus with internal RFM type variables, provides the key to finding these valuable but hidden customers.