Using Machine Learning to Identify ‘Never Pay’ Subscribers

You go to a phone store, pick out a handset, the salesperson checks your credit history, then 15 minutes later you walk out with a new phone – and the mobile operator has no way of being certain that you will ever pay a bill.

40 percent of mobile phone operator bad debt – billions of dollars a year – is due to subscriber fraud or default. Fraudsters use real, stolen, or fake credentials to acquire significantly subsidized smartphones with no intention to pay the monthly subscription fees. Others start off innocently enough, but within 18 months or less will stop paying.

Credit risk and subscriber fraud are key concerns for global carriers. In markets that enjoy high revenues per user (RPU), where carriers heavily subsidize handsets, telcos assume a lot of risk because they lack information about the subscriber’s ability or intention to pay.

Even credit rating agencies acknowledge that existing systems need to be bolstered in order to contain the damage from ‘never pay’ or high risk subscribers. Most operators rely upon solutions that are neither systematic nor automatic in the way they check subscribers. Some depend on the availability of a just a handful of fraud analysts. In some geographies, all that stands between a good subscriber and a lost handset are phone store employees checking a spreadsheet.

What if there was a system that could accurately detect a high-risk subscriber? What if it could predict who would default – ninety days before they first fail to pay?

The field of machine learning predictive analytics is showing enormous promise. Machine learning analytics is increasingly viewed as a key tool to improve the efficiency and profitability of mobile carrier networks. Industry research firm International Data Corporation (IDC) forecasts worldwide spending on cognitive and artificial intelligence (AI) systems will increase by 59.3 percent to USD12.5 billion by the end of 2017, and will reach USD46 billion in 2020.

In trials and deployments analyzing real carrier data, new big data/machine learning algorithms and neural network architectures have successful enabled operators to accurately predict subscribers’ intention and ability to pay monthly service bills by analyzing massive amounts of data. They draw on a variety of demographic, social, CDR and financial data sources to provide a 360-degree analysis of subscription applications.

This technology approach not only identifies undesirable subscribers at the time of signing on for a mobile service, but also successfully predicts delinquencies in the following 60 – 90 days. And the use case can be extended beyond subscriber fraud prevention. The same predictive analytics approach can be used to generate credit profiles for the unbanked, whose credit-worthiness cannot be assessed using traditional systems. This has huge implications for both developed and developing countries and has the potential to redefine the cost model for subscriber credit checking.

The original version of this article was published on the blog of Argyle Data. It has been reproduced with their permission.

Vikash Varma
Vikash Varma
Vikash Varma is the President and CEO of Argyle Data. He brings 23 years of multi-disciplined experience to the business, having established a successful track record of building businesses from the early stages to market leadership.   Argyle Data develops machine learning analytics tools for mobile operators, enabling detection, prediction and prevention of fraud, security and revenue assurance issues.