EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

352. You Shall Not Pass - Minimizing false-positive ensemble classification through threshold optimization

Invited abstract in session MD-57: Methodology in asset allocation and banking, stream Modern Decision Making in Finance and Insurance.

Monday, 14:30-16:00
Room: S06 (building: 101)

Authors (first author is the speaker)

1. Richard Oberdieck
Banking Circle ApS
2. Ruben Menke
Banking Circle ApS
3. Christian Møller Karsten
Advanced Analytics, Banking Circle

Abstract

As a bank, we handle millions of transactions a day. Almost all of these (>99.99%) are completely fine and legitimate transactions. However, every once in a while a sanctioned entity is either trying to send or receive money with a payment we process. We want, and are legally required, to stop such payments. Due to the high volume of transactions a day, algorithms play a crucial role in flagging potentially suspicious payments. If any of these algorithms judge there to be a match to a sanctioned entry, the payment is investigated manually.

This poses a challenge though: the number of true positives is very low, so even the best algorithms will mostly flag false positives (>95%). This means almost all of the payments an analyst goes through are fine, they just have an unlucky similarity to a sanctioned entity. To reduce this issue, the algorithmic thresholds have to be as tight as possible, while still finding all the true positives.

In our presentation, we will show how this problem can be posed as a mixed-integer programming problem and the results it yields. We also discuss how to extend the formulation to allow for a risk-based approach, where a certain number of misses of true positives are allowed, in order to further reduce the number of false positives. This results in a fascinating pareto front, which is used as a basis of discussion around price of catching all true positives.

Keywords

Status: accepted

Back to the list of papers