Machine Learning in Customer Life Cycle: Insights, Explorations and Predictions

customersA short article to illustrate machine learning (ML) benefits in a marketing campaign to reduce the customer churn rate. The scenario is rather simple with customer generated data of a telco provider. Data are basically made of the customer monthly charges for every specific time frame of the day as well as information about the customer itself, its type of subscription and its relation with the provider (historical use of services like the hotline for example). One campaign is organized in order to propose discounts for the next 6 months to selected customers, hopefully the ones that are planning to stop the subscription.

And this is where the machine learning model that predicts probabilities of leaving customers helps not only to propose a list of selected customers, but also to give insights about the best discount value given the campaign budget and the targeted ROI.

Lets have a look at the following figures giving results with a rather bad predictor using logistic regression (on the left hand-side) and an excellent one using random forest classifier (on the right hand-side). The models are trained using historical customer data (previous period). 5 different discount values (5%-40%) and customer take rates are simulated (32.7%-63.3%). The first two figures correspond to the campaign costs including 6 months of charges for leaving customers that didn’t accept the offer or where not targeted by the campaign. The next two figures are the ROI using again 6 months life time value. The last two figures give the performance curves (False/True positives, False/True negatives) as a function of the probability threshold used in the decision process of the ML algorithm. A small threshold value leads to a large number of customers selected for the campaign.

There are five interesting observations.

  1. The optimal discount value varies with the quality of the prediction. A good prediction using advanced ML model allows to increase the proposed discount and therefore the take rate.
  2. The threshold (controlling the number of customers in the campaign) that minimizes the costs also depends on the accuracy of the model. Threshold value needs to be adapted to the accuracy of the prediction.
  3. The accuracy of the model has, as expected a tremendous impact on the ROI.
  4. Rather bad predictors can also bring valuable returns if we choose the right parameter values (threshold, discount).

So the use of a machine learning prediction to tune marketing campaigns targeting a reduction of the customer churn rate is an excellent way of exploiting large historical data sets especially when campaign parameters are tuned correctly such that the first generation model is a success and gives benefits right away. Of course, next generation models will become more accurate as we get new experience, better know-how and more data, such that campaign ROI really takes off.

Indeed, once historical data are collected and available, machine learning is rather easy to develop and is fast enough to explore and simulate many scenarios before engaging campaign investments.

Data science is about insights, exploration and doing accurate predictions, all three together.