There are a number of controls and tools that leading CSPs are putting in place to reduce financial losses from subscription fraud, the topic of my last two posts.
Before getting started, CSPs should determine who “own”’ and is accountable for subscription fraud. Is it the fraud team? Credit risk? Revenue assurance? Also, is there a clear and agreed fraud risk appetite that has exec sponsorship and is agreed by all stakeholders?
Another major factor to consider is the business model of the fraudster. Is it feasible to break the business model and take the value out of committing the fraud instead of, or as well as, investing in fancy solutions that promise to help prevent fraud? In the case of subscription fraud where the device is the primary driver for resale (often on the international stage), is there a way to render the device of little to no value on the resale market?
There are numerous examples of the industry coming together to share data of known frauds and profiles in order to mitigate the impact of fraud. Data sharing for these purposes should be non-competitive: fraud is everybody’s problem, but still relies on a common definition of what data to share, where to securely host the data, how regularly it should be updated and what mechanisms are available to access the consortium data.
Once all of these questions have clear answers, then analytical techniques can dramatically help reduce subscription fraud losses. These include machine learning models and social network link analysis.
At the pre-book stage, there is a shift from the more traditional predictive models using regression techniques toward machine learning models. In part, this is due to the ever-changing nature of fraud. Whilst the scorecard approach is tried and tested, the time required to develop the score means by the time it is operationalised the fraud MO has changed, leaving the scorecard less effective than it could be.
On the flip side, machine learning techniques are scalable and can be self learning and adaptive, so they keep up more effectively with changes in tack by the fraudsters. Machine learning increases the concentration of fraud relative to non-fraud applications at high score thresholds, whilst minimising false positives and therefore impact to customer experience. At the lower score thresholds machine learning has the opposite effect, reducing the concentration of fraud relative to non-fraud, therefore minimising false negatives and fraud losses. This is because algorithms such as neural networks are able to understand hidden layers within the relationships between variables contained within the data, and so make more accurate decisions.
Early Warning Signs
Once activated, the early usage of a new subscription can be monitored to detect warning signals. For instance, if the subscription has zero usage within the first few days after an order is fulfilled, does that indicate unusual behaviour?
Early warning alerts have traditionally been deployed by rules-based systems, which are effective to a point. However, with the ever-changing nature of fraud, dependency on rules introduces a latency in deployment, and it can be difficult to manage rulesets that may quickly become obsolete. As with the pre-book stage, this is where adaptive machine learning techniques can be deployed in order to identify changing patterns and effectively prioritise alerts for fraud analysts to review. With the complexity of fraud, and a general approach of having to do more with the same resources, machine learning capability enables fraud teams to make efficient use of analysts’ time in order to protect the business and minimise any service impacts to genuine customers.
Other analytical technology can further enable fraud teams to become increasingly proactive at either pre- or post-book stage. Link analysis software using advanced algorithms can piece together hidden relationships in the data to quickly identify fraud rings. Uncovering one fraud case can lead to finding many more cases from the same fraud ring by leveraging organised gangs’ biggest weakness – shared identity data – using fuzzy matching and relationship-driven predictive analytics. This allows fraud teams to achieve much higher detection rates and lower false positives. Visualisation techniques allow investigators to significantly reduce the time taken to unpick fraud rings and identify greater volumes of likely fraud in order to reduce potential losses.
Follow my posts on this blog for more information on telecommunications fraud.