Faced with fighting an ongoing pricing war that impacts heavily on their bottom line, South African cellular networks need to explore alternative methods of retaining subscribers.

Increasing revenue by poaching customers

Over the last few years, it has become apparent that the South African cellular market is approaching maturity. Rather than being able to rely on an ever-expanding consumer base, South African network operators are realising that, for all intents and purposes, the market has reached saturation. The simple reality of this situation is that the only way in which a network will be able to substantially expand its specific customer base is by "poaching" those customers from one of their rivals.

The networks themselves would have been among the first to realise this when looking at their subscriber base, and they have responded in a typical manner; the lowering of prices which has resulted in a pricing war. This is an understandable reaction and it is effective up until a point.

Cell C recently cut its prepaid rate to 66c/minute with impressive results; it was able to increase its subscriber base by approximately 3 million users between December 2013 and April 2014, with the majority of those users coming from rival networks. Now Vodacom has recently slashed its prices, with MTN following suit.

Floating subscribers

However, in reality, the majority of clients won't leave their network, even if the call rate is dropped substantially. That being said, there are a substantial number of "floating subscribers" who will move between networks based on the best call rate. The problem the networks have is any reduction in call rate has to be across its whole subscriber base. This means there must be a point where the addition of new subscribers at substantially reduced rates results in a negative return. However, if your rivals cut their rates, there is often no choice but to follow suit.

An alternative suggested here is to concentrate on retaining the current subscriber base. One way in which the networks could prevent subscribers from "churning" (in the sense that they move across to a rival network) would be to offer their existing subscribers special offers and reductions. If they offered these across their entire subscriber base, they would essentially be faced with the same problem of negative returns. However, if there were some way of targeting floating or at-risk subscribers with these special offers, then the possibility exists that networks could retain these subscribers without losing money in the process. This is a straight-forward calculation; if the cost to the networks of the reduced offers is less than the expected future spend of the at-risk subscribers, then it becomes economically viable.

Thus, there are two steps to this process:

1. Identify at-risk subscribers.

2. Calculate the expected future spend of those subscribers and evaluate it against the marketing spend required to retain them.

The second step relies primarily on spend history data, which all of the networks will have. It also requires the networks to take a view on the probability of the at-risk subscriber actually being retained as a result of the special offer, along with the additional length of time that the subscriber will stay on the books. Both of these can be refined over the course of time.

Predictive analytics and churn modelling

Therefore, the tricky part of the whole process is the first bit; identifying the at-risk subscribers. The only reliable way in which this can be done is to utilise predictive analytics, and in particular, churn modelling. Simply put, predictive analytics allows the user to predict future outcomes based on past experiences. This is done by developing a mathematical model which is essentially a simplification of reality as pertains to the situation being modelled. These models are built by looking at large data sets representing past behaviour, and representing that behaviour through mathematical algorithms. Churn models are a specific application of predictive analytics, which predict the likelihood of a subscriber churning based on past behaviour.

In this case, the churn model would look at the behaviour of all of the subscribers, past and present. The data would be broken up into "churners" and "non-churners". Then, it would analyse the behaviour of all the churners up until the month they left and look for patterns. At the same time, it would also look for patterns in behaviour for all of the non-churners. For example, perhaps churners had a particular pattern of decreasing call use, or maybe they had increased dropped calls in the months leading up to the churn. The more information that can be used to build the model, the more effective it generally becomes. Once built, the model should be able to differentiate between churners and non-churners, and attach a number between 0 and 1 (known as propensity) to the likelihood of a subscriber churning.

Continuing with a simple example, if the model has been built purely on the behaviour described above, it should be able to then look at a random subscriber's spend pattern and number of dropped calls, and be able to say with a certain propensity if that subscriber will be a churner or a non-churner. At the end of this modelling process, the result will be a list of subscribers along with their propensity to churn. The network can now design a marketing campaign bearing in mind the expected future spend of the subscribers. This is also a rich area for analytics, as optimisation models can be built in order to maximise the outcome of this process. However, this is not the focus of this discussion.

Consequences for the SA networks

The networks are at the point now where any further reduction in price is going to have serious consequences on their profitability, but the reality is it may not be enough to retain customers.

Churn modelling provides a viable alternative to reduce losses or increase revenue in a mature market, and is a route which bears serious consideration.

In practice, the implementation of an effective churn model can be fairly complicated as it requires the integration of different data sources and rigorous data cleaning before any modelling work can commence. The churn solution at its core is a predictive model that is heavily reliant on the incorporation of in-depth, business understanding in the form of business rules. These rules are invariably unique considerations, specific to each company, that need to be incorporated into the predictive process. In addition, issues like gaining an acceptable confidence level in model output and the treatment of the uncertainty associated with such models, is crucial to obtaining meaningful results.