WeirdFlows: Anomaly Detection in Financial Transaction Flows

WeirdFlows: A Novel Approach to Detecting Anomalies in Financial Transaction Flows

Introduction

Financial crime, especially fraud and money laundering, has become increasingly complex with the rise of digital transactions and international financial networks. Traditional investigative methods struggle to keep up with the volume and sophistication of illicit activities. The WeirdFlows pipeline, introduced by researchers from Italian universities and the Anti Financial Crime Digital Hub, presents a novel method that leverages network analysis to detect suspicious transaction flows without relying on predefined patterns or labeled training data. This approach focuses on interpretability and scalability, essential for aiding Anti-Financial Crime (AFC) analysts in formal investigations.

The Challenge of Financial Crime Detection

Financial markets involve a multitude of actors and transactions that form complex, dynamic networks. Fraudulent schemes often employ intermediaries and convoluted transaction paths to mask illicit activity, making detection difficult. Existing methods either require prior knowledge of fraud patterns or fail to provide sufficient explanation for detected anomalies, limiting their practical utility in AFC operations.

Network analysis has gained traction as a valuable tool in this domain, modeling transactions as weighted, directed temporal networks. Various techniques have been developed, including community detection, social network metrics, graph learning, and machine learning models. However, many suffer from interpretability issues or dependency on annotated datasets.

Bastian Schwind-Wagner_Follow "WeirdFlows offers a powerful network-based method for detecting anomalous financial transaction flows without relying on preset fraud patterns. Its interpretability and scalability make it an effective tool for uncovering complex money laundering schemes and sanction evasions in large-scale banking data."

WeirdFlows Pipeline: Methodology and Implementation

WeirdFlows addresses these challenges by introducing a top-down search pipeline that explores all possible transaction paths up to a maximum length between nodes in a financial network. Each node can represent entities at various granularities such as countries (ISO codes), banks (BIC codes), or individual accounts (IBANs). The pipeline consists of building weighted temporal transaction graphs aggregated over chosen time intervals (weekly or monthly) and applying a recursive depth-first search algorithm to identify all possible paths starting from a selected node.

The key innovation lies in defining the transaction flow weight between two nodes as the sum of the minimum edge weights across all possible paths connecting them through intermediaries. This model captures attempts to hide large transfers behind multiple smaller transactions routed through various intermediaries.

WeirdFlows also integrates time series analysis on the weights of detected flows, using moving averages and forecasting models to identify significant deviations indicative of anomalous activity. This temporal dimension enhances the ability to spot emerging fraudulent behaviors that may not be reflected in static network snapshots.

Application and Results

The WeirdFlows pipeline was evaluated on a large dataset of 80 million cross-border transactions over 15 months from Intesa Sanpaolo, Italy’s largest bank. The focus was on detecting attempts to circumvent economic sanctions imposed by the EU following the outbreak of war in Ukraine in February 2022.

Analysis at multiple aggregation levels revealed that while direct transaction volumes between certain country nodes remained stable, indirect flows through intermediaries exhibited abnormal increases after sanctions were enacted. For example, flows routed through an anonymized intermediary country showed a 66% increase in transaction volume compared to expected patterns.

By isolating these intermediaries and analyzing their transaction flows, AFC analysts could identify suspicious actors and complex laundering schemes hidden beneath apparently normal transaction patterns. The interpretability of WeirdFlows’ results allowed for transparent investigation steps supported by quantitative evidence.

Advantages of WeirdFlows

WeirdFlows stands out for its ability to:

• Detect complex transaction flows without needing predefined fraud patterns or labeled data;
• Scale effectively to handle tens of millions of transactions across multiple temporal granularities;
• Provide interpretable, explainable results that support formal AFC investigations; and
• Identify anomalies in transaction flows by integrating network topology with time series anomaly detection.

These features make WeirdFlows a promising tool for financial institutions and regulators aiming to enhance fraud detection capabilities amid evolving financial crime tactics.

Conclusion

The WeirdFlows pipeline represents a significant step forward in anti-financial crime technology by combining network analysis with temporal anomaly detection in a scalable and interpretable framework. Its successful application to real-world banking data demonstrates potential for broader adoption in monitoring complex financial systems and improving the detection of illicit transaction patterns, particularly in challenging contexts such as sanction evasion.

Dive deeper

• Research ¦ Arthur Capozzi, Salvatore Vilella, Dario Moncalvo, Marco Fornasiero, Valeria Ricci, Silvia Ronchiadin, Giancarlo Ruffo; “FlowSeries: Anomaly Detection in Financial Transaction Flows?”. arXiv:2503.15896. doi: 10.48550/arXiv.2503.15896 ¦ Link