Churn: Predict vs Pre-empt

Customer churn is quite topical for companies with subscription based offerings. Apart from the old adage around the costs of customer acquisition vs retention, many tech companies not only have a dependency on scale for profitability, they see the inverse of marginal returns with incremental subscription. While acquisition initiatives are often obvious, retention has historically been taken for granted. Churn, defined as “number of individuals moving out of a collective group”, is a challenge and rife in industries with subscription services, like mobile telecoms. Churn prediction has thus gained in popularity over the years. A lot of analytical resources is directed toward churn propensity modelling and there’s been much progress made in exploiting the results. I refer to ths conceptually & focus on alternate approaches to the subject as a whole.

Broadly, regression or classification tree techniques are applied on a few variables, some number crunching takes place usually resulting in a propensity score by subscriber. These techniques perform exceptionally well on two main dimensions:

1) Duration, 2) Consumption.

In summary, one expects, the likelihood of churn to decrease as customers increase their usage on the number and value of services, over time. Modelled in a contract cycle, the churn risk on your post-paid base, is expected to be higher during the sunset period (i.e. as the contract nears its end). While in the prepaid base, although unrestrained, the opposite is usually true. One can expect loyalty over time. (Illustrated right):

While the approach to customer retention should be seen holistically, if you are considering or busy with propensity modelling in the manner described above, let’s first start with the implications of the ‘subscription economy’. Engagement or disengagement no longer follows a smooth profile. Analysing consumption data alone, as the leading indicator, is onerous. What are the causals?

I like to think of the concept as one would operate a gas fuelled burner. You'd notice in turning of the flame, one would tends to turn of the valve at the gas bottle first. This allows the gas in the pipes to drain through. Finally the tap at the cooker is shut off. Now looking at the flames alone, one couldn’t be blamed for suggesting the dying flames was a gradual process. In reality, the point of termination (turning of the valve) was an instantaneous action.

For the mobile subscriber look first for triggers then trends. Spend & consumption patterns can have trends (illustrated on the right), but when extrapolated is this a prediction?

A drop in usage, or change in behaviour oftentimes follows a trigger. These rank among the top:

• A drop in network quality
• Better price form competitor
• or a bad experience at one of the contact points.

But this is not unknown. In defence the call for 360 degree customer data has been going on for some time now. Historically, the emphasis in ERP systems has been accounting rather than activity. Finding accurate customer activity repositories is rare. Yet we can, given knowledge of the above triggers, still do more than calculate propensity scores. Should you invest in hindsight intelligence or accurate & timely service level feedback?

More often than not, service level reports exist. These metrics spark a tactical conversation. Possibly because, unlike our machine counterparts, we thrive on aggregates. Actions on aggregate information, seldom meet the real time needs of our value seeking customer. But we are living in a time when data has moved from main frames to the cloud. Analytical software is available freely and code can be reused with minor adjustments. We thus have opportunities to move from static estimations, to programmed activities and finally machine learning. Statistical process modelling, illustrated below, opens prospects for live engagement. We can ping and reach out to customers, proactively, with a whole host of customized messages, not just on name, but purpose. Live data begins with collation, then analysis and finally action.