Public Patrick Papsdorf Adviser European Central Bank Discussion

Public Patrick Papsdorf Adviser European Central Bank Discussion: “Detection and Explanation of Anomalous Payment Behavior in RTGS Systems” by Trieples, Daniels, Heijmans 15 th Payment System Simulator Seminar 31 st August 2017, Helsinki/Finland The views expressed here are those of the author and do not necessarily represent the views of the European Central Bank and the Eurosystem.

Rubric See https: //jakubmarian. com/wp-content/uploads/2014/12/milk-consumption. jpg See https: //jakubmarian. com/wp-content/uploads/2017/03/nuts 2 -researchers. jpg P. Papsdorf - Discussion anomaly detection 2 www. ecb. europa. eu ©

Rubric Summary Anomalous payment behaviour Anomaly detection in payments data • Demand: high for identifying anomalous behaviours timely, (semi) automatically with high accuracy • Challenges: data complexity (3 V’s), networks, scope setting, resources … • Paper: Idea • Apply machine learning to help identifying anomalies P. Papsdorf - Discussion anomaly detection Aim Result • Can method identify liquidity problems of a bank in the data? See Practical Machine Learning: A New Look at Anomaly Detection by E. Friedman, Ted Dunning • “Method worked reasonably well. ” 3 www. ecb. europa. eu ©

Rubric Summary AI / machine learning Anomaly detection (outlier detection) Identify items/events/observations that do not conform to an expected pattern or other items in a dataset. See https: //betanews. com/2016/12/12/ deep-learning-vs-machine-learning/ Here: Unsupervised anomaly detection. Detects anomalies in unlabelled data sets. “You don’t know exactly what you are looking for. ” See Practical Machine Learning: A New Look at Anomaly Detection by E. Friedman, Ted Dunning Supervised anomaly detection based on labelling of data as "normal" and "abnormal". Autoencoder • Feed-forward neural network, which learns from examples. It applies learnings to new data. No learning of concrete examples but recognition of patterns. See Wikipedia on neural networks • Trained to reconstruct input layer at the output layer by processing input via a hidden layer Hidden Output layer in which a set of neurons form a compressed. Input representation of the layer input in a lower dimensional space. (See Triepels, Daniels, Heijmans) See https: //blog. keras. io/building-autoencoders-in-keras. html P. Papsdorf - Discussion anomaly detection 4 www. ecb. europa. eu ©

Rubric Summary [Please select] What was done here Dutch component of TARGET 2, customer payments of 20 most active banks broken down in time intervals and liquidity vectors Training dataset • Nov 08 – Aug 09 Holdout dataset • Sept – Oct 08 Test data • LEARNING • OPTIMISATION • Sept – Oct 09 • EVALUATION • Three different autoencoders • Optimal point of neurons/compression determined: more neurons would not lead to much better reconstruction (low MRE), i. o. w. dynamics of liquidity vectors well captured. • Anomalies spotted (above set threshold) and three observation areas (A, B, C) examined by drill down using time interval, banks, in/outflows • Bank that was subject to bank run was identified P. Papsdorf - Discussion anomaly detection 5 www. ecb. europa. eu ©

Rubric Comments and questions • Thank you for this! • Novel approach to apply AI to payments data. • Autoencoder method resulted in identifying outliers. • Possibly opening a new strand in FMI analytics interesting for System Operator and Oversight. • Many potential fields could be considered, like AML/CTF, fraud, intraday liquidity management, funding issues, MM outliers, autotriggers/alerts, interdependencies … P. Papsdorf - Discussion anomaly detection 6 www. ecb. europa. eu ©

Rubric Comments and Questions Model set-up • One hour time intervals for liq. vectors vs longer/shorter intervals. • Reasoning for chosen timespans of datasets. • How to operationalize method in a dynamic environment continued learning? Anomaly detection • Threshold setting and review is manual. • Type 1 and Type 2 errors to underpin “reasonable accuracy”. • Compression level: risk of Over-fitting vs. Under-fitting. • Timeliness and accuracy – some outliers (A, C) only explained as non-relevant over time. P. Papsdorf - Discussion anomaly detection 7 www. ecb. europa. eu ©

Rubric Comments and Questions Scenario • Chosen scenario typically evolves very differently. • Once run is detected it may be too late. • Are there other earlier signals in payments data? E. g. CB operations, interbank, delays, intraday credit usage, cash reservations that could be tested. • Knowledge of bank run helped to understand identified outlier (B). P. Papsdorf - Discussion anomaly detection 8 www. ecb. europa. eu ©

Rubric Thank you … and a discussion appetizer for later L. Page: “Artificial intelligence would be the ultimate version of Google. ” Ethical and moral aspects S. Hawking: “The development of full artificial intelligence could spell the end of human race. ” related to AI E. Musk: “biggest risk we face as a civilization. ” V. Rometty: “this technology will enhance us. So instead of artificial intelligence, I think we'll augment our intelligence. ” P. Papsdorf - Discussion anomaly detection 9 www. ecb. europa. eu ©