V1. If you were running an e-commerce business for eggs, what variables would affect your churn model and why? 2. Give us the mathematical equation of the churn model with the variables you identified above.
Answers
Explanation:
You could build this, and it'd probably be pretty interesting, but it wouldn't be very useful*. Let's say I join your service and, after a week, your model tells you that I have a 95% probability of churning out today. What are you going to do about that? Are you going to send me an email to try to keep me? Offer me a discount? How do you know those actions will work?
Rather than predicting the probability of some event happening or not (which you don't really need to predict, because you can just wait for it to happen and track the event itself, not the prediction), you'd be better served by measuring which features impact that event the most. Now let's say I joined your service from finding it on Google organically. Maybe you send me some emails during my tenure on your service that are different from a control group to whom you send emails. Maybe you show me a product catalogue that's different from the product catalogue you've shown to other users that joined via Google, just for kicks. Maybe you experiment with me by not showing me ads, by showing me more ads than you show other users, or by showing me different types of ads. And now after two weeks of being your guinea pig, I decide that I don't need your service anymore and leave it for good. But you've collected some data about how long I retained, how much money I spent, how many friends I brought into your service in that time, and you've also collected that data about the people that joined your service via organic search on my first day, too. Now you can compare my results with theirs using some pretty simple statistical methods.
You've got something actionable in the second scenario: you know that an independent variable (like acquisition source) has been shown to affect a dependent variable (retention) in a certain way. You also know that you can logistically regress my behavior for one independent variable, holding other independent variables at the mean value, and predict an odds ratio of a dependent variable having a certain value. You can also feed by behavior into a model that considers certain actions independent of other actions and predict whether or not some result (like monetizing) is likely. These are all much more actionable than the model which predicts (probably incorrectly) how likely I am to churn out today.
*I have found one use for the "will X user churn out today?" model: I call it the Hail Mary discount, but I think it probably only works in mobile freemium gaming. If a user is likely to churn out today, I will offer him a treasure trove of in-game items at an extreme discount in the hopes that I can monetize him. But this isn't optimal, because hopefully I've been optimizing that player's experience from the very beginning of his use of the service.
1. The variables that would affect an e-commerce business churn model are:
- Average subscription length: It is the amount of time an average customer spends paying for a company’s goods or services.
- Customer acquisition cost: It is the amount of money spent to acquire a single new customer. A high customer acquisition cost will create a need to maintain a high customer retention rate for the efficient growth of the company.
- Customer lifetime value (CLV): It measures a customer’s total value to a business over the duration of their relationship.
2. The mathematical equation of the churn model is given below:
- E-commerce customer churn represents the number of customers whose relationship with the company has come to an end.
- The churn rate is calculated by dividing the number of customers lost or churned by the total number of customers.
- Listening to customers and understanding what they desire can help the company reduce churn.
#SPJ3