Anti Money Laundering Automation
with predictive modeling
ABN AMRO BANK
Anti Money Laundering
• What are Anti Money Laundering activities?
• Who defines the policies and regulations?
• Who implements the monitoring and reporting?
• What are the technical challenges?
• Millions of transactions per day.
• A rule-based monitoring system gets filtering rules from business lines and
processes the transactions and raises alerts.
• The raised alerts are put in to the backlog to be manually processed.
• A large part of the raised alerts are false cases that are dismissed as not
needing further investigation.
• Predictive modelling comes into play to help make a second layer
automated filtering to throw away the false alarms and mitigate the
problem of workload.
1. Information regarding the transaction such as amount, date and time,
origin and transfer path, target account.
2. Information regarding the alert coming from the rule-based system.
3. History of the transactions from and to the target account and other
accounts of the client.
4. History of the past alerts and analysis reports by the Customer Due
Dilligence experts, made on the target account or other accounts of the
Model (feature engineering)
• We create features from the
– Generated alert,
– Target transaction
– Interbank transfer route
– History of the account
– History of the client
– History of the past raised alerts on the account and the client
– Client’s profile, etc.
• We take binary targets based on the analysis records:
– Non-suspicious dismissed alerts (negative class)
– Alerts that were filed for further investigation (positive class)
• An ensemble (e.g. random forest, gradient boosting) model is trained on
all the alert level features and targets.
Model (tuning approach)
• Set classification threshold in a way to avoid any false negative: most
conservative approach to make sure no actual suspicious activity passes
• On average we are able to cut down the false positive alerts of the rule-
• The model is re-trained periodically with new analysis records
• A small set is always analysed to make sure the model performance stays
• On workload: less manual time and energy spent on dismissing false
alerts, more time to filter and investigate cases
• On the economy: Anti Money Laundering (AML) regulations in general
contribute to transparent financial flow in the Netherlands, and EU.
Predictive modelling helps banks achieve this with less cost and more
• On society: transparent financial flow helps closing loopholes and prevent
financial corruptions. By reducing the possibility for financial criminal
activities, it can help make national and EU financial policies more
effective, improve the mutual trust individual and the society have with
the banks, and strengthen the economy.
– AML cash: deposits and withdrawals
– AML LRT (Large Reportable Transactions): wire transfers
– AML HRG (High Risk Group): cross-border transactions from/to outside the
• FCCM – a tool that rule based selects suspicious transactions
• CDD – Client Due Diligence, they receive alerts in the backlog and
manually process the alerts for further investigation.