Financial Crime News and Fraud Risk Management (AI and Machine Learning in Banking)
Attention fraud fighters! For all the latest news around AI in risk management, AI and machine learning in banking, anti money laundering for banks, and fraud risk management, please subscribe: https://www.youtube.com/@Feedzai.RiskOps CHAPTERS: 0:00 Using AI to Improve AML Transaction Monitoring 0:59 Detection: A 3-Step Process 3:20 Operations: Helping Analysts Spot Suspicious Patterns ABBREVIATED TRANSCRIPT: AI has been in the news a lot recently. All the image generation tools like Dall-E and Midjourney and now ChatGPT and even the upgraded Bing, the search engine. At Feedzai, we are big on AI – it has always been part of our DNA. I want to talk to you about how we use AI to improve the experience of AML transaction monitoring systems. Hi, this is João and this is Feedzai Financial Crime News Weekly Update. In an AML transaction monitoring solution, there are multiple places where we can make AI work in favor of the financial institution. So you can split into two big areas. On one hand: detection. You want to make the regulators happy, you want to reduce the risk of your financial institution being exposed to financial crime. And on the second hand, you have more of the operational side. Compliance analysts opening SARs and all of this, and AI can have a very big impact in these two areas in different ways. So for the detection, we advise our customers to go into a three-step process. You start with rules. You can use your solution's rules or your existing ruleset, and then rules will generate alerts. And alerts will generate activity from your operations team, so it will be SARs or escalations. With this, you can then start your journey to a supervised machine learning. So you'll train supervised models based on these feedback signals. And after you get confidence on your supervised models, you can start by reducing the false positives because the model will learn patterns that are around the rules. So say the rules are catching these things that you can define really well, the models will actually see more of the pattern in a more holistic way and enable you to get more confidence on the very, very low confidence rule trigger. So even if rule number one or rule number two trigger, the model will say, “No, no, we've seen this in the past year or two, this is with 99% certainty a false positive.” And after you get that confidence, now you can actually use the model to reduce the alerts from the rules, but you now have here an opportunity. So with less alerts coming from the rules, you can now use the model itself to generate more alerts. The third stage is introducing unsupervised models, which will be there to really go after the known unknowns. There's a lot of activity that we know it looks suspicious, but we don't even know what to look for. We cannot think of a rule that will enable us to uncover that. Imagine if I want every week to have ten alerts of very strange activity. I don't know if it's going to be suspicious or not -- it's just strange. The model has never seen this in the past. That's what the model would trigger. And if it triggered, you generate alerts. Generating alerts will allow your operations team to review them. You will have the feedback, and now you can incorporate this back either into your rules or your model -- your supervised model. So at the end of the day, you'll end up with a very mature system that will combine rules -- so let's say static AI -- supervised learning, and unsupervised learning, all in one system that will help you be more agile in detecting AML. The second area where AI can help is when we look at the operational side of transaction monitoring. And here it's all about how can we make the analyst spend time in what's more important for them. So the typical grunt work, we want to make it as automatic as possible and the human work needs to be where the juice is at. Typically analysts working lots of systems. They need to copy paste information from here and there, put it all in a big Excel spreadsheet, and then do some manual queries in that. And what I mean by making this easier is that we already have all that in one place, right? You have the transaction monitoring system, you have some reference data and some enrichments over there. If it's all in one place, we should be able to aggregate all that information, and when an alert triggers, present that information in a way that is useful for the analyst. Instead of going manually over hundreds of transactions of some alert, we should be able to have the analysts look at this in a more efficient way, typically through visual aids. It's no longer about spending time putting information from one place to another and then manually going over it, but it's more about spending time looking at the pattern that is in front of you in your computer and making the assessment if it is suspicious or not, or if it's worth an escalation or eventually if you need to open a SAR.
Attention fraud fighters! For all the latest news around AI in risk management, AI and machine learning in banking, anti money laundering for banks, and fraud risk management, please subscribe: https://www.youtube.com/@Feedzai.RiskOps CHAPTERS: 0:00 Using AI to Improve AML Transaction Monitoring 0:59 Detection: A 3-Step Process 3:20 Operations: Helping Analysts Spot Suspicious Patterns ABBREVIATED TRANSCRIPT: AI has been in the news a lot recently. All the image generation tools like Dall-E and Midjourney and now ChatGPT and even the upgraded Bing, the search engine. At Feedzai, we are big on AI – it has always been part of our DNA. I want to talk to you about how we use AI to improve the experience of AML transaction monitoring systems. Hi, this is João and this is Feedzai Financial Crime News Weekly Update. In an AML transaction monitoring solution, there are multiple places where we can make AI work in favor of the financial institution. So you can split into two big areas. On one hand: detection. You want to make the regulators happy, you want to reduce the risk of your financial institution being exposed to financial crime. And on the second hand, you have more of the operational side. Compliance analysts opening SARs and all of this, and AI can have a very big impact in these two areas in different ways. So for the detection, we advise our customers to go into a three-step process. You start with rules. You can use your solution's rules or your existing ruleset, and then rules will generate alerts. And alerts will generate activity from your operations team, so it will be SARs or escalations. With this, you can then start your journey to a supervised machine learning. So you'll train supervised models based on these feedback signals. And after you get confidence on your supervised models, you can start by reducing the false positives because the model will learn patterns that are around the rules. So say the rules are catching these things that you can define really well, the models will actually see more of the pattern in a more holistic way and enable you to get more confidence on the very, very low confidence rule trigger. So even if rule number one or rule number two trigger, the model will say, “No, no, we've seen this in the past year or two, this is with 99% certainty a false positive.” And after you get that confidence, now you can actually use the model to reduce the alerts from the rules, but you now have here an opportunity. So with less alerts coming from the rules, you can now use the model itself to generate more alerts. The third stage is introducing unsupervised models, which will be there to really go after the known unknowns. There's a lot of activity that we know it looks suspicious, but we don't even know what to look for. We cannot think of a rule that will enable us to uncover that. Imagine if I want every week to have ten alerts of very strange activity. I don't know if it's going to be suspicious or not -- it's just strange. The model has never seen this in the past. That's what the model would trigger. And if it triggered, you generate alerts. Generating alerts will allow your operations team to review them. You will have the feedback, and now you can incorporate this back either into your rules or your model -- your supervised model. So at the end of the day, you'll end up with a very mature system that will combine rules -- so let's say static AI -- supervised learning, and unsupervised learning, all in one system that will help you be more agile in detecting AML. The second area where AI can help is when we look at the operational side of transaction monitoring. And here it's all about how can we make the analyst spend time in what's more important for them. So the typical grunt work, we want to make it as automatic as possible and the human work needs to be where the juice is at. Typically analysts working lots of systems. They need to copy paste information from here and there, put it all in a big Excel spreadsheet, and then do some manual queries in that. And what I mean by making this easier is that we already have all that in one place, right? You have the transaction monitoring system, you have some reference data and some enrichments over there. If it's all in one place, we should be able to aggregate all that information, and when an alert triggers, present that information in a way that is useful for the analyst. Instead of going manually over hundreds of transactions of some alert, we should be able to have the analysts look at this in a more efficient way, typically through visual aids. It's no longer about spending time putting information from one place to another and then manually going over it, but it's more about spending time looking at the pattern that is in front of you in your computer and making the assessment if it is suspicious or not, or if it's worth an escalation or eventually if you need to open a SAR.