Keeping your existing customers engaged is vital. Promotions can make a significant difference in retention, what offers should you send to customers and who should receive the offers?

You might start by segmenting your customers into groups and comparing their behaviour. If young, female customers have a higher chance to churn you could target these customers with higher offers.

This kind of analysis seems simple and intuitive, but it becomes much fuzzier when you have multiple offers and multiple customer groups (i.e. what should we prioritise age groups or gender if there are conflicting trends)?

Ibex utilises a number of machine learning algorithms that can solve these problems.

Treatment model

The first step is predicting, for each customer, the likelihood of depositing if they receive an offer. Ibex is using deep learning models to make these predictions, but a simple regression model is probably a good start. Such a model will predict a percentage for each customer and each offer, and allows you to forecast what will happen if you send offer A vs offer B.

Baseline model

But there could be a chance that we'd be better off not to send anything at all to some customers, for example if they are already highly engaged. In this case, the customer is likely to visit your site regardless - and a baseline model can capture this effect.

It is similar to the treatment model, in that it gives a probability of deposit for each customer, but assumes no offer has been sent.

Combined with the treatment offer this can be used to predict the impact of each offer, i.e. how do we expect the offer to change this customer's behaviour?

Model selection

At Ibex, we have spent years tweaking and optimising the machine learning models used for these prediction tasks.

Since the models are the basis for decisions regarding which offer to send to which customer, it is important that these model are well tested and reliable, and that all required information is fed into them so that the decision can be as informed as possible.

‍

‍