Definitions – Predictive Churn and CLV scoring

The concept of churn in the retail industry is tricky since often there is no definite action that terminates a relationship, but rather a decline in frequency until the frequency is really low. Our model predicts this decline in frequency and present a likely timeline where normal frequency is still upheld by the customer.

Even though the predictive churn and CLV scoring is built on complex data modelling, the output is easily accessible and actionable in Engage.



The predictive model is essentially a churn scoring, trying to estimate which customer that is likely to churn, but includes several calculated values as well, explained below.

The likelihood for churn is presented as a number between 0 and 1. The closer to 0 a customer is, the less likely the customer is to churn. From this value we also calculate other useful predictive values.

Definitions of the predictive scorings

  • Churn score and activity level
    A value based on customer history (purchases, open rates, clicks etc) and the results from our AI modelA value between 0 and 1 is calculated, indicating the likelihood for churn. From these scorings we also calculate customer activity level.

    0.0–0.5 = Active (Low churn score)
    0.51–0.75 = Declining (Medium churn score)
    0.76–1.00 = Leaving (High churn score)

  • Act before
    This is the date that the model predicts as the last date before the member has churned. After this date you can expect the customer as lost.
  • Average days between purchase
    The average amount of days between purchases. The average is calculated from the first purchase date to the lastIf the member has less than two purchases, this value is 0. The reason for this being that with only one purchase no number of days has passed between two purchases. The frequency is yet unknown.
  • Years remaining as active customer
    The estimated number of years left as active customer.

  • Customer Lifetime Value
    The estimated value for the remaining customer relationship. In this context it’s not a value to mix up with the total customer lifetime. The CLV is calculated from a fixed manually entered margin and a cost of capital rate. Default values are 25 % and 2 %.

  • Purchased recently
    The alternatives are simply YES/NO. This score is based on the number of days between purchases, and we continuously check if thcustomer currently is within their normal purchase cycle or not. The score is calculated on the average days between purchases + one standard deviation.

    Remember—the scoring is based on individual data and calculated on customers normal behavior. A member with relatively low recency can still be scored as “YES”, since we know that this is her normal behavior. And at the same time a member with relatively high recency can still be scored “NO” since the model has noticed a dip in her purchase pattern.

    A customer with only one purchase will be attributed a score based on the member base average until the point that she has sufficient data for an individual prediction.
Was this article helpful?
2 out of 2 found this helpful


1 comment

  • Hi, 
    For the 'avg days between purchases' why do you use always the first purchase and the last? Would it not make more sense to do last purchase minus the latest purchase before that? Instead of always using the first purchase? Because then for a frequent customer the avg dys between purchase would still increase and look like many days, vs one that for e.g. has made two purchases but just a week a part.


Article is closed for comments.