Izengard Analytics and Models - Key Detection Model Differentiators
In many cases a range of statistical inference models are needed (much like the data mining days) and you don’t have to reach out to machine learning. Many of these statistical models are effective at catching fraud (for example Bayesian techniques) and do not take up the compute power and resources needed for machine learning. Izengard data scientists are conscious of when to use techniques that won’t cause a large usage overhead for our clients
Izengard employs state of the art in-memory anomaly detection which is scaled at minimum for 3 years of data but can be scaled for more. In any case aggregates for further years, suspicious events history and payee, value, volume history is always kept. Izengard brings best practices that is across all 3 domains of Cyber-security, IT Risk Management and Financial Crime and therefore does not need multiple anomaly detection models which are consuming the same data and more or less doing similar detections. Izengard’s anomaly detection approach improves the quality of detection and makes it ready for risk scoring. In addition, Izengard costs institutions less than multiple siloed solutions.
Many of Izengard’s typologies for fraud and money laundering have been built using supervised methods. The applicability of these are then trained on client data and where necessary augmented by our consortium data. As indicated Izengard’s models are adaptive, so they will learn and respond to new signals sometimes with Human in the Loop. Izengard is a believer in Causal AI and use these techniques in combination with known techniques to deliver you better detection of typologies. Typologies discovered through unsupervised techniques can be promoted to supervised techniques once the risk indicators and flow of funds in the typology are known.
Izengard takes the notion that in effect, typologies are viruses that mutate. As they mutate there are new challenges in discovery and re-inforcement learning allows the Izengard environment to react to it with the basis of cumulative award being accurate detection.
In addition to this, we have seen mistakes in traditional machine learning to misinterpret correlation as causation. This approach is wrong as items can be positively or negatively correlated. Traditional complex machine learning models also do not fully uncover explainability with complex methods, our partner uses Causal AI to overcome this and we co-work with our client SMEs to create more robust AML/Fraud models.
Izengard combines the above techniques and uses Causal AI consciously in order to determine fraud, bribery, corruption or money laundering. In general one technique alone is too limited to achieve stronger detection, so Izengard uses ensemble models. Overall, since Izengard is using singular models with scale and performance, they are cheaper than operating multiple siloed models and have a higher accuracy ratio of prediction of a suspicious event, due to the multiple risk factors from the 3 domains taken into account.