Our approach deeply uses machine learning techniques in order to detect any financial crime right away. The model uses historic data to learn how to spot risky or abnormal behavior exhibited by transactions, clients, suppliers, or other players. It uses two types of models:
- Supervised learning algorithms, that tell us how similar to past fraud a new transaction is and
- Unsupervised learning algorithms, that tell us how odd a new transaction seems when compared to past transactions.
The first model guarantees accuracy, the second the ability to adapt to changing realities. You can find many financial crime solutions in the market suffering from two problems:
- many false positives and
- long investigation times.
Therefore, our approach propose machine learning models to solve both consecutively.
Fixing Many False Positives ;
We propose using machine learning for optimally combining existing or new rules into rich fraud indicators that, based on tried and tested maths, ensure you are way more likely to get relevant alerts in a much smaller sample of investigation efforts. Our machine learning models have a supervised and an unsupervised component. Supervised machine learning models focus on distinguishing within historic data known past fraud cases from the remainder, becoming really good at with the KPIs, spotting future fraud that behaves like past fraud. Then, we can evaluate the quality of the model on a separate set of data. And deploy it to real time or in batch. Which is great, but not enough. In fact, fraudsters are creative people, if one approach to fraud does not work, they will try another and another until one does. So our financial crime detection approach needs to be able to accommodate surprises. We do this by using unsupervised models. This type of models focusses on profiling typical past transactions and against these spotting odd ones. Not necessarily fraudulent. But odd; therefore, it is worth to investigate.
The great part of solution is that both models can be created by your business users and field experts, who inevitably understand fraud KPIs better than anyone. The solution provides a familiar interface and a safe zone where business users can:
- test existing KPIs,
- evaluate their respective fraud predictive ability,
- create new richer KPIs to monitor any sort of fraud intervenient (from customers to branch employees or suppliers) for any type of financial crime,
- create the machine learning models that optimally combine existing KPIs into rich fraud indicators,
- evaluate model worth on an impartial sample of data,
- deploy the models in batch or to real-time,
- set thresholds only for the result of the models,
- and do what-if analysis to estimate the investigative burden the tresholds represent.
Solving the problem of long investigation times
Risky transactions will be investigated by operators, who must decide for each transaction whether it is criminal or not. Such work consists of many tasks and data from several data sources having transaction’s history and intervenient from all disparate sources.
More importantly, as transactions get investigated and a conclusion is made regarding whether they were actually fraud or not, this information is used to monitor model health over time. If the proportion of false positives at a point starts increasing, this is a sure sign that it is time to update the supervised and unsupervised models to incorporate the latest environment changes. A warning can be sent when model update is advised. Or model update can be called automatically or at predefined schedules.