Abnormal Traffic Pattern Detection in Real-Time Financial Transactions

EasyChair Preprint 827

Abstract

We have developed a combined statistical analytical, machine learning (ML) and deep learning (DL) approach to detect abnormal traffic patterns in financial messages involving monetary payment instructions. We used optimally anonymized historical transaction data from multiple financial institutions from disparate geographic locations globally. Our objectives were to provide client institutions with customizable levels of alert notification based upon their risk tolerance, and the ability to detect and prevent fraudulent payment instructions in real time. Our statistical analytical approach demonstrates that a preliminary transaction-based calendar can be established based solely on historical transaction data containing message counts and their arrival times, and can be further improved based upon user input as necessary. Several ML and DL models were built and evaluated for each of their performance metrics (e.g., accuracy, confusion matrix). Our results suggest that a time series ML model (seasonal autoregressive integrated moving average (SARIMA)), and particularly two DL classification models (Autoencoder and Restricted Boltzmann Machine (RBM)) can consistently yield highly accurate predictions. Our study also suggests that ML and DL models in conjunction with a statistical analytical approach provide a powerful tool for real-time anomaly detection in financial transactions.

Keyphrases: Financial fraud, anomaly detection, deep learning, financial transaction, fraud detection, machine learning, statistical analysis

@booklet{EasyChair:827,
  author    = {Sean Rastatter and Travis Moe and Amitava Gangopadhyay and Alfred Weaver},
  title     = {Abnormal Traffic Pattern Detection in Real-Time Financial Transactions},
  howpublished = {EasyChair Preprint 827},
  year      = {EasyChair, 2019}}