Traditional financial crime (FC) systems take a blunt approach to identifying FC, similar to the ‘Stop and Search’ approach sometimes used in law enforcement. However, much more precise methods exist. In a similar way to sniffer dogs, machine learning can help to ‘sniff out’ FC with much greater accuracy and efficiency than traditional approaches.
In 2016, 32.5% of Stop and Searches in London were considered to have a positive outcome (i.e. stopping the suspect was considered ‘worthwhile’) and 20% led to an arrest.* Proportionally, FC systems are even less effective. Typically, 3-5% of alerts are considered worthwhile and only 1-2% lead to action being taken – e.g. a Suspicious Activity Report (SAR) being disclosed.
By contrast to the Stop and Search approach, sniffer dogs can have upwards of 85% accuracy.** Machine learning models (MLMs) can drive similar improvements within FC.
Like sniffer dogs, MLMs are able to more effectively detect criminal behaviour. For example, they:
- Learn without having to be taught explicitly how to identify criminal behaviour: They are presented with a range of different situations and taught the desired outcomes in those situations. They are not taught how to determine these outcomes but instead learn for themselves how to do so. For example, a MLM may be provided with a range of customer records, their transactions and whether or not a SAR was raised against each of them. It can then be trained to determine which customer behaviours led to SARs being raised. As with sniffer dogs, once MLMs have been trained using situations where the outcomes are known, they can make predictions in situations where the outcomes are not known - e.g. customers which have not been analysed for suspicious behaviour
- Identify subtleties humans cannot: Rules implemented in FC detection systems are simplifications of behaviour considered to represent FC, generally based upon a limited set of data. MLMs can create much more accurate representations of criminal behaviour by considering a broader range of data, allowing them to identify behaviour which humans cannot
- Continue to learn over time: Sniffer dogs are rewarded for accurately identifying behaviour of interest. Similarly, the creation of a feedback loop in MLMs means that they can continue to learn and adapt to emerging FC typologies.
Whilst MLMs offer many benefits, caution must be given to promises of near 100% accuracy. They can be taught to predict outcomes very accurately using training data but too precise results are likely to mean that they will be unable to generalise their ‘understanding’ when used in practice.
MLMs also offers similar benefits to sniffer dogs when compared to traditional approaches:
- Lower operational cost: More accurate detection of criminal behaviour means considerably fewer false positives, fewer investigators and lower operational cost
- Better identification of criminal behaviour: Modelling criminal behaviour more accurately provides the opportunity to identify criminal activity which is currently being overlooked
- Reduced risk of reputational damage: Blunt rules can lead to poor decisions – e.g. exiting customers based on spurious links to terrorism. Well-trained MLMs can help reduce the risk of inappropriate generalisations being made
- Better customer experience: Customer experience can be significantly disrupted by inadequate FC systems - e.g. customer payments being delayed. MLMs can help to ensure customers are only investigated where there are reasonable grounds for suspicion.
Better understanding how MLMs differ from traditional techniques, and the benefits they offer, will enable FC professionals to take a much more intelligent approach to FC risk mitigation.
**Their accuracy is dependent on the application and the environment in which they are being employed. See Jezierski, T. et al, (2014) Efficacy of drug detection by fully-trained police dogs varies by breed, training level, type of drug and search environment, Forensic Science International, Vol. 237, pp 112-118.