•
•
Resulting scenarios are often not efficient or effective in identifying suspicious activity and subjective threshold setting is difficult to defend to regulators
Heuristic or rules based approaches rely on factory settings or expert input in order select risk factors and set thresholds
Heuristic or rules based
This is an improvement, however scenarios may still be sub-optimal and may not consider a broad range of inputs or more nuanced interactions
Analytically informed approaches consider patterns of suspicious activity to raise or lower thresholds to iteratively optimize performance
Analytically informed
•
•
•
•
This affords much more freedom to improve performance, considering a wide range of factors and portfolio specific drivers to optimize performance
Statistical and quantitative analysis leverages statistical techniques to identify correlations and clusters of activity that can be used to help identify suspicious activity
Statistically optimized
Machine learning
•
•
They can “learn” over time based on feedback on their own effectiveness – identified false positives lead to improvements in future performance
Machine learning tools leverage prescribed inputs and outputs to build sophisticated algorithms (e.g. GBM, deep neural nets) that identify suspicious activity based on multiple data inputs and complex behavioral patterns
Current state of the industry