February 24, 2020
February 24, 2020
by Ugo Salvati | 6 min read
As the telecom industry reached its saturation point, the telecom giants have shifted their focus from designing marketing offers, aimed at acquiring new users only, to campaigns aimed at reducing the churn rates of existing subscribers. Spending hundreds of euros on marketing to acquire new users, seems now less relevant than overlooking those users switching to a different operator. Seeing an existing user go, in facts, results not just in the loss of the initial marketing investment to acquire it but also implies giving up all the future revenues expected from such user. Not only, retaining loyal users also provides telecom operators with the opportunity to up-sell and cross-sell new products and services to them.
Our Customer, by the time of our engagement, had already attempted to implement its user retention strategy by building a predictive Machine Learning (ML) churn model. Our data science team assisted in recommending the most suitable Artificial Intelligence (AI) and ML models with the goal to better understand existing users’ preferences and, subsequently, reduce their churn rate. KDM FORCE, then, deployed predictive churn modelling skills to forecast the probable future users’ behavior, this in order to identify the churn-prone profiles.
Our data science team leveraged the Customer’s Big Data architecture to access extensive data (though anonymised), train the ML model and design a user-friendly interface, working hand in hand with the Customer, to align a prototype’s features with the Customer’s needs. This all before putting the model into production, after a successful completion of a series of tests.
At the end of our project, we also assured that knowledge would be transferred to the Customer providing the skills needed to keep enhancing our predictive models or build new ones, without external help in the future.
Each step described in the above Customer Case follows KDM FORCE’s proven analytics approach, as this is synthetically defined in the image here below.
As outlined, firstly, we define success criteria in close cooperation with our Customer. For telecoms this is typically represented by an increase in the ARPU (Average Revenue Per User).
We then apply a well-tried approach for data mining the available data. In the above Customer case: the voice and data traffic consumption. This has the objective to detect changes in usage patterns and user experience.
In the third step, we, typically, begin analysing and understanding users’ preferences by building a model. Our model separates False Positives (users who would not switch to a different operator despite their behavior) from False Negatives (those who, undetected, would switch) and then penalises False Negatives on the assumption that it is far more costly to win a new user than to retain an existing one. Our approach is also enriched by other factors, for example, historic data on churned users from previous years. Seasonality for one. We have often observed that most dropouts concentrate in a specific time of the year while, in other moments of the year, we consistently see high retention rates. The bottom line is that availability of good quality data, covering at least one year, is essential for training good ML models.
In the next step of our approach, we release the offline prototype for testing purposes and illustrating functionality and extract business value.
As a fifth step, the resulting models are deployed on production systems, providing our Customer with intuitive dashboards and graphic interfaces that provides with the ability to forecast which users are most likely to churn and offering them only relevant personalized bundles.
Together with the delivery of a playbook to keep the models updated, the final and crucial step of our way to go leverages a long-term strategy for upgrading the ML models and the overall predictive user churn approach. This to help our Customer achieve a double-digit reduction in their user churn strategy.