Prerequisite: In order to use the churn prediction model, your Cleeng integration must generate engagement data through the Cleeng entitlement engine. Read more.
What is the churn prediction model?
Cleeng’s churn prediction model assigns a (1) churn probability score and (2) a churn reason to your subscribers. This is done using a machine learning algorithm, which learns what customer churn looks like in your service.
How can I use these predictions?
There are 3 main ways you can use these predictions:
- Identify which of your subscribers are at risk of churn - so you can automatically target them
- Identify why your customers are likely to churn - so that you can create the right retention message
- Simplify your customer analysis by directly seeing which customers are loyal - and which are wavering
The algorithm allows you to predict user loyalty, better understand why customers consider cancelling, and target those customers so that you can retain more revenue.
What are some examples of how I can use predictions?
These concepts are explained in more detail below, but the most typical scenario is a customer having high churn probability due to low content engagement. You may have experienced this with your own subscriptions.
These customers consider cancelling because they are conscious of how little they use the service. Netflix or Amazon Prime use such insights to trigger two type of retention tactics:
- Highlight upcoming content to the customer
- Offer the customer a significant discount
If the customer's payment price is reduced from $5.99 to $2.99 (for example), they will often consider the lower price worth paying simply because there might be something they would like to watch in the future. After all, they signed up to your platform in the first place for that same reason.
Rescuing a relatively small percentage of such subscriptions every month generates significant revenue over time.
So what does the model predict?
The model makes three key predictions:
- The probability of a customer churning before their next renewal
- The reason why at-risk customers are likely to churn
- The total revenue that is highly likely to churn
Every subscriber who meets the model’s conditions will be assigned a churn probability score. If that score is under 50%, the customer is not identified as being at risk. If it is above 50%, they are.
However, the model also distinguishes between customers more or less likely to churn. So you will see your at-risk customers split between those who show moderate risk, and those who are extremely likely to leave your service.
The second prediction made by the model shows why this customer is likely to churn. The model uses 5 primary reasons:
They are not engaged with content
This means that the customer has a low level of engagement, that is similar to other customers of yours who have cancelled. 'Low engagement' is defined purely in the context of your customer base.
They are burning through content
This means that the customer is engaged at a very high level, but in a pattern that resembles other highly engaged customers who quickly churned. This is the 'burn and churn' phenomenon. Customers binge watch a catalogue or series, and then cancel.
Their payment is not reliable
This refers to both the customer's current payment method, and their payment history. the combination of the two allows the model to learn patterns that lead to 'involuntary' churn.
They have a history of churning
This relates to the customer's previous subscriptions to your service. It looks at both the number of previous subscriptions, and the duration of those subscriptions compared to the current one.
Their offer is problematic
This means that the offer itself is the main reason the customer is going to churn. It refers to any unique attributes that the offer has, like the content it grants access to, the channel where it's available, or the price.
The algorithm makes sense of these reasons within your customer base. So there is no standard measure of ‘low engagement’, for example. Instead, the model is saying that a customer’s engagement pattern looks a lot like the engagement patterns of other customers who recently churned from your platform.
How does the model treat voluntary and involuntary churn?
This is an important distinction. Here is the best way to understand what the model is identifying in each of these cases:
When the algorithm is predicting that a customer will churn due to payment reasons, the event it is predicting is the termination of the subscription. This is slightly different to voluntary churn cases.
Voluntary churn occurs when a customer cancels their subscription. This stops the subscription renewal process, and concludes with the termination of the subscription at the expiry date.
The prediction model identifies these two phases (cancellation and termination) separately. It first predicts the risk of a customer cancelling their subscription. When a customer has cancelled their subscription, it will predict their risk of terminating their subscription. This is normally very high, unless you are consistently recovering a lot of cancelled subscriptions.
How does the model operate?
The model continuously reviews the last 2 weeks of activity on your platform, and will adjust its predictions to match what it is seeing. A subscriber’s churn risk is identified once the following conditions are satisfied:
- The model has seen enough subscribers churn from your service in the past 2 weeks to understand what churn looks like on your platform (typically 100 cases)
- You currently have at least 100 active subscribers
- The subscriber has a monthly subscription (the model is most accurate for monthly cycles)*
- The subscriber has been active for long enough for their engagement patterns to be meaningful
*The only exception to the above is when a customer cancels their subscription renewal. In this case, the subscriber is automatically classified as ‘very high risk’ no matter what their subscription cycle is (eg. annual).
How is churn risk identified?
The model uses a wide array of payment, subscription, and engagement variables to identify churn patterns on your platform. For example, it may recognise that a customer’s engagement level is declining in a manner very similar to customers who have churned in the past.
It will also look at factors like how long they have been a subscriber, the offer they are subscribed to, whether they have a history of churn and so on.
Furthermore, the model continually learns based on changing patterns in your subscriber base. For example, a ‘risky’ engagement level may change based on seasonality, and the model will recognise that the patterns that lead to churn are changing as well.
How accurate is the prediction model?
There are different ways to understand the accuracy of a churn prediction model.
- How well does the model predict churners and non-churners on my platform?
This is the accuracy measure companies with machine learning products communicate publicly. Using this measure, the Cleeng churn prediction model averages around 95%.
But this definition does not fully reflect the model's predictive power. As most of your customer's don't churn, a better measure is one that pays more attention to false positives (predicted to churn but didn't) and false negatives (predicted to renew but churned).
For this an ROC curve is the best way to truly measure how well a model predicts outcomes. Using this measure the model averages 87%. This is classed as 'excellent' in machine learning. However, many Cleeng broadcasters have scores of over 90%, which is classed as 'outstanding'.
What analytics can I use?
Your churn prediction analytics are available on the Retain dashboard:
- Number of subscribers at low / moderate / high / very-high risk of churn
- Reasons why customers are at risk of churn
These analytics are a nice shortcut for subscriber data analysis, as they will let you know who is happy with your service, who is wavering, and who is about to leave. Anticipating your subscriber churn allows you to plan accordingly.
How to target at-risk customers?
You can target at-risk customers by setting up segment actions. You just need two attributes to create the trigger for your marketing campaigns:
- Churn risk
- Churn reason
So for example, you can create a segment that is customers with an 80% or higher churn probability where the primary reason is disengagement.
Once this is in place, you no longer need to take any action in the Cleeng platform. Customers who meet these conditions will automatically be pushed to whatever campaign you use on your marketing platform.