Using Machine Learning Analytics to Detect Fraud Scams

Electronic payments have paved the way for new and old authorised push payment (APP) scams to hit the UK market, with a 40% rise in 2019; 56% of Europeans surveyed experienced at least one type of fraud/scam in 2018/2019. Fraud scams are a concern for every fraud manager.

In my last post, I discussed three best practices for fighting scams. Now let’s look at how analytics can support you in protecting your customers.

The Data Challenge

As a fraud manager, you typically rely on categorised, tagged data and historical trends for the identification of fraud trends and model training. Detecting fraud without marked data can be a challenge. As scams are constantly changing, a method to identify a scam based purely on historical data would be ineffective, so you need to find another way.

FICO uses patented techniques to reduce the need for this data, which is a big step change for the industry. Self-calibrating behavioural analytics technology is the way to go.

What Is Self-Calibration?

The idea behind self-calibration is the model’s ability to adjust and adapt to distribution changes of predictive variables, to match the evolving nature of payment patterns

Note: SL and SR are percentile estimates, which are derived and updated in real-time

Figure 1 illustrates how abnormal behaviour is deduced for any given feature. By computing the variable distribution (in real time), the analytics are able to identify normal behaviour. Consequently, any behaviours outside of these normal patterns will be identified. Furthermore, by dynamically scaling variables and estimating the distribution, the analytics can reassess what is deemed normal behaviour for any given time period. For example, money transfers during December would potentially be higher than November, given the holiday spending period.

FICO uses self-calibrating analytics to understand all aspects of the payment behaviour transaction history and anomalous behaviours based on the personalised transaction behaviour of the accountholder. This assists in making a more refined decision on the status of the transaction.

Why does this matter to you?  Self-calibrating analytics are ideal to use in situations where there is insufficient data on which to build historical information, which is a reality when it comes to scams.

Multi-Layered Self-Calibrating Profiles (MLSCP)

Multi-layered self-calibrating (MLSC) technology produces a score indicative of fraud or genuine based on a multitude of specialised latent behavioural transaction features. The underlying concern with MLSC models is to determine unusual transactions, based on the quantiles of key features.

The architecture of an MLSC model (Figure 2) resembles that of a neural network; however, the hidden layers are the self-calibrating models of latent features as detailed above. As data is input, these models continually calculate the distribution of each feature for each group (as defined by the different colours in the input layer — these groups are defined to minimise mutual information overlap in feature categories used in the model). Within the hidden nodes, no group is contained within each node more than once, which reduces multi-collinearity and instability.

The reasons behind this are twofold; firstly, no hidden layer has an overdependency on any particular group, thus reducing selection bias. Secondly, as each node connects different features in multiple ways, we are able to get differing perspectives of the data through a collection of unsupervised devised latent features. The output of the model is a score between 1 and 999.

What makes MLSC models so flexible is that, upon collection of any tagged historic data, the model can be tuned to account for these new data points; this enhances model performance by tuning the contribution of the unsupervised latent features on the output later of the network. As such, MLSC models are a semi-supervised algorithm and are ideal for human/AI cooperation, particularly on new problem areas where data is lacking, non-existent or rapidly developing and changing.

Learn More

With scams there will be little to no tagged data, and your fraud management program needs to ability to detect anomalous behaviour of your customers to shut the fraudster down. With the use of Multi-Layered Self-Calibrating Profiles, you will have the ability to identify the outlying behaviour on your customers in real-time, allowing you time to reach out to your customer to confirm the transaction.

For more information on this technology, refer to the our post Fraud Analytics for Open Banking: Multi-Layered Self-Calibrating Models. And see this executive brief on how to combat fraud scams with the FICO Falcon Platform.