We are going to describe hereafter what is predictive analytics? How it benefits traditional business lines in many areas and how these technologies were applied successfully to predict bad-debt in OTT services offered by Swisscom.
What is predictive analytics?
Predictive analytics is not about a crystal ball that can come up with a lottery number or be able to know if it is going to rain tomorrow or not.
It is rather a branch of advanced analytics using statistical and machine learning methods to learn from the past by building models trained on historical transactional data. The subsequent models after training are then evaluated for its precision and optimised to deliver accurate predictions.
Predictive analytics can do so by finding patterns, many times not obvious or understandable to us, or by finding relationships within the data that will be indicate likelihood of a particular class, risk measurement or forecasted value.
Of course, as opposed to descriptive technologies, normally used by traditional business or business intelligence departments, it will pose an extra effort to mine available data and extract all its hidden value.
We sometimes refer to those techniques as data mining. To mine is a good analogy to the fact that if we want to find the gold nuggets hidden in the mountain of data, we will need to add some extra effort and make use of modern technologies.
Nowadays, most of the current business use cases that could benefit from predictive analytics technologies don’t necessarily need a large amount of data to start obtaining those real benefits.
Predictive analytics will provide benefits from the very beginning, and it does so, by helping to identify many business data issues like data quality, duplicates, missing data, non-unique id’s and so on. But not only that, implementing data mining techniques, will also help us in better understanding customer profiles, behaviours and possible wrong doing or fraud. Hence, we are extracting value and insights when mining raw data.
Optimization of business processes: Forecasting bad-debt in OTT services
At Swisscom, we know from accumulated experience that many traditional business processes can be optimized very rapidly and will start improving business KPI’s when adopting these modern predictive analytics strategies. Optimization starts from day one!
We can easily set up similar predictive use cases by using another related use case, which have been developed previously, in an end-to-end fashion. The reason for this depends not only on the similarity of the use case but also on the data understanding, exploration and preliminary mining activities needed.
We will now describe in more detail our latest predictive case that was successfully put into production mode very recently and that evaluates the risk of bad-debt in a continuous mode of operation. 
Swisscom Direct Carrier Billing
Swisscom direct carrier billing, NATEL® Pay, is a service offered to Swisscom customers having a mobile subscription or prepaid number where they can perform electronic purchases to third party merchants by means of their mobile telephones. Purchases like, apps, games, music, films and tickets, executed in major online stores with this mobile payment method will be directly billed to the Swisscom customer or deducted from their current pre-payment balance. Hence, there is no need of credit card and additional fees or introducing payment details each time and registering in multiple sites. The whole mobile payment solution centralizes the payment and billing processes of different merchants.
Predictive Analytics use case
This payment method service is being used the more and more by Swisscom customers as it provides a simplified and secured means of payment. Customers also benefit from a transparent and centralized method that facilitates online purchases.
Among all the customers, a tiny amount of them, won’t pay their invoices for those OTT services and hence will be set as bad-debt. Even though they represent a very small number of cases, the amount due from those purchases is not negligible.
The predictive analytics use case was then well defined together with business partners and aimed to identify, as soon as possible, those customers that won’t be likely to pay their invoices. The system will be able to raise warning flags to operational departments for their consideration.
Data and Business Understanding
Once we have precisely defined the use case, we can start collecting business and historical data that will provide insights about what is going to be forecasted. Those datasets often consist of raw data that is not exploitable in that unprocessed form.
Predictive analytics starts hence with a business understanding phase that goes along with the data understanding counterpart. It allows both data scientists and business partners to better understand the business use case, requirements, data availability, operations needed as well as any missing data, problems with quality or database issues regarding well defined keys.
Within those phases, the prediction’s data granularity and variables to be used for building the predictive models are also defined.
The business case being introduced here is of additional difficulty as it is analogous to finding a needle in a haystack. Very small number of customers won’t pay their bills, so special attention must be paid when identifying bad-debt and not to cause any inconvenience to good customers.
Traditional vs Modern technologies
Until very recently the OTT department was entirely relying on traditional risk services. Those traditional (rule-based) services, to evaluate risk of bad-debt, base their risk assessments on previous historical information from customers about payment/debt on other similar services.
Typical problems experienced by rule-based approach are related to not being able to establish predictions for new customers or simply missing/inaccurate/incomplete data is being used for those customers.
Besides, rule-based methods don’t generalize well, predictive performance tends to be poor and methods provide static predictions based on previous business experience, which at the same time, don’t adapt dynamically to changes.
Additionally, business partners don’t have the means to evaluate forecasting performances: many times, bad-debt risk decisions are being provided by those external services. Furthermore, relying on those risk services also makes everything like a black-box without any control on it.
On the other hand, we can find modern predictive technologies, which provide the obvious benefit of being able to identify bad-debt customers with high accuracy and diminishing their associated losses. Additionally, we can also highlight some of the associated rewards that are often overlooked.
- Models continuously adapt to data, i.e.: they learn from the past and adapt dynamically to changes. When models deviate much from expected errors, they can automatically be triggered for re-training
- Models normally provide a relative importance measurement of the variables used for the prediction, hence, allowing much better understanding of forecasts
- They can reveal data insights totally unattended or hidden to business experts
- They can be applied to similar use cases in the business line without much effort
- Can also reveal profound insights about customers behaviour and of fraudsters
- Customer insights arising from analytics methods can be used to evaluate current health of business and its evolution along time, helping business to define new strategies
- Predictive insights can be adopted on marketing campaigns or be displayed in the form of traditional business intelligence dashboards
- Model performance can be enhanced along time by adding more explanatory predictive variables incrementally
End-to-end forecasting machinery deployment:
Prototyping and creating a first model that can predict bad-debt with high accuracy is only the first step of setting up and activate the predictive analytics machinery in modern business.
Successfully putting in place all the business processes and components needed for the automation of daily forecasts in terms of bad-debt identification, requires a great deal of expertise and knowledge of different technologies. It is still rare among data scientists the ability to master end-to-end knowledge of needed skills. Because of this, many predictive use cases never manage to run on a productive environment.
As data scientists, we need to have a diverse knowledge of distinct technologies so that we can make use of what is already in place to avoid, for instance, data replications or benefit from database computing capabilities, etc. It is important to understand, in broad terms, the existing technologies so that we can better design the predictive solution and avoid any issues when running in a productive environment.
In this particular predictive case, we made use of Teradata and SAP databases technologies for distinct reasons.
Teradata is where our business data resides. SAP HANA is where our analytical dataset had to be built and models to be trained for predictions. We also made use of low level predictive analytics capabilities available in SAP HANA databases which require trained and expert knowledge. In addition, the daily batch scheduling of the bad-debt predictions was implemented in SAP BW/4HANA databases. Operational departments are in this way able to monitor normal running of model’s predictions.
The forecasting capabilities have been deployed end-to-end in a production mode very recently and OTT business units are already taking advantage of those.
In parallel, we built additional business intelligence services in form of dashboards for the OTT business unit. They help business very much when visualizing customer segments and behaviours.
Finally, we are also delivering general insights from our pre-built predictive datasets that help to improve business figures and KPI’s. Those insights apart from providing very accurate regular snapshots of current customer’s behaviour and preferences, will also be of excellent value when approaching new clients or promoting new services.
Rethink your business for the Digital Revolution:
Let us help you transform your traditional business processes with predictive analytics solutions.
If you have already envisaged some predictive use cases or you want support in brainstorming them, you can contact us directly by email:
Matthias Mohler (firstname.lastname@example.org) – Head of Analytics
Sergio Jimenez (email@example.com) – Senior Analytics Consultant
 Swisscom: How Is Switzerland’s Largest Telco Company Managing Risk as a Payment Provider?
 Swisscom (Schweiz) AG: Managing Risk with Predictive Analytics