FlowSeries: flow analysis on financial networks

Abstract

Abstract The digitalization and automation of anti-financial crime (AFC) investigations has made significant progress in recent years. However, key challenges remain—in particular, the need for interpretability in the output of AI models and the limited availability of labeled data for training. Criminal activity in transaction networks often involves complex, evolving patterns specifically designed to evade detection. We introduce FlowSeries, a top-down flow analysis methodology to explore transaction data and analyze complex interaction patterns over time. Rather than relying on pre-defined patterns or labeled training data, our approach scales to large transaction volumes and provides interpretable insights into anomalous behaviors, aiding AFC analysts in their investigations. We evaluate the effectiveness of this method using a dataset provided by the bank Intesa Sanpaolo (ISP), comprising 80 million cross-border transactions over a 15-month period. In collaboration with ISP’s AFC experts, our analysis focuses on detecting anomalous transactions and identifying suspicious actors in the context of the economic sanctions imposed on Russia following its invasion of Ukraine on February 24th, 2022.

Keywords

• Network analysis
• Anti financial crime
• Anti money laundering
• Temporal networks
• Graph search algorithm