Interesting, Miles. I’ve wondered for a while how many FIUs (or banks) that are relying on these models are building them, basically, inductively or deductively. Are they using known cases to build a dataset of ML patterns and then looking for those. Or are they building models of Not-AML and then looking for aberrations? Any idea?
Thanks for reading! And the Bluesky conversation on this was fascinating. It’s definitely a topic which deserves more research. Generally I think there are enormous inefficiencies in how generic typologies trickle down to TM alert thresholds which are often poorly specified.
Interesting, Miles. I’ve wondered for a while how many FIUs (or banks) that are relying on these models are building them, basically, inductively or deductively. Are they using known cases to build a dataset of ML patterns and then looking for those. Or are they building models of Not-AML and then looking for aberrations? Any idea?
Thanks for reading! And the Bluesky conversation on this was fascinating. It’s definitely a topic which deserves more research. Generally I think there are enormous inefficiencies in how generic typologies trickle down to TM alert thresholds which are often poorly specified.