Suspicious bank transactions detected better with data analysis

Machine learning, cluster analysis, and anomaly detection ensure faster and more effective identification of suspicious financial transactions.

What’s the problem?

Like all other members of the sector, this international Dutch bank is making major efforts to comply with the Act to Prevent Money Laundering and Financing of Terrorism [Wet voorkoming van witwassen en financieren van terrorisme]. Unusual transactions must be reported to the Financial Intelligence Unit of the Dutch Ministry of Finance. Non-compliance can result in a high penalty being imposed. The bank detects these transactions with a computer system that applies criteria based on predetermined rules. Those criteria are derived from previous suspicious practices, such as high transaction amounts. Bank staff then check the unusual transactions. This approach is in fact inefficient and expensive, because it produces a lot of “bycatch”: out of 100 transactions noted as unusual, 99 were found to be harmless. That low score is thought to lead to staff becoming less alert, resulting in a greater risk of errors.

Suspicious bank transactions detected better with data analysis

So what’s the solution?

The basis for a solution was data analysis by PwC of almost 2.5 million transactions. We did this with a start-up that had developed an advanced analytical technique for this purpose. Using the results, our data analysts constructed a two-part system. First of all, cluster analysis was used to divide the group of individual customers up into several subgroups with a specific behavioural pattern. That pattern consisted of dozens of dimensions, such as the size of transaction amounts, cash card withdrawals abroad, number of transactions, timing of transactions, etc. Deviations from that specific behaviour trigger an alarm. Machine learning causes the group structure to change dynamically with each individual transaction. PwC also developed a mathematical system for anomaly detection: which transactions differ from the previously determined transaction behaviour?

And what’s the result?

The bank now needs to do much less work to detect suspicious transactions. Cluster analysis and anomaly detection immediately reduced the number of transactions that were wrongly considered to be unusual from 99 percent to 50 percent. Achieving a score of twenty percent is possible once final improvements have been made. For customers of the bank, that means that their transactions are less likely to be blocked unnecessarily. Also, the more specific behavioural criteria increase the likelihood of a truly suspicious transaction being noticed. The system is therefore much more efficient, with a greater likelihood of identifying terrorist financing and money laundering.

Contact us

Casper Ruizendaal

Partner, PwC Netherlands

Tel: +31 (0)62 227 45 09

Follow us