WeirdFlows: Anomaly Detection in Financial Transaction Flows

Arthur Capozzi, Salvatore Vilella, Dario Moncalvo, Marco Fornasiero, Valeria Ricci, Silvia Ronchiadin, Giancarlo Ruffo · 2025

Read WeirdFlows: Anomaly Detection in Financial Transaction Flows by Arthur Capozzi, Salvatore Vilella, Dario Moncalvo, Marco Fornasiero, Valeria Ricci, Silvia Ronchiadin, Giancarlo Ruffo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15896.txt.

Introduction

Arthur Capozzi and six co-authors present WeirdFlows, a system for spotting fraud in financial transaction networks, in a two thousand twenty-five paper. Anti-financial-crime investigators face unlabeled data and models they cannot explain, while fraud patterns keep shifting.

WeirdFlows searches a transaction network top-down for anomalous flows and actors. It needs no preset pattern library and no training set, and it explains the anomalies it flags so analysts can act. The team evaluated it on eighty million cross-country transactions from Intesa Sanpaolo spanning fifteen months. Bank experts confirmed its effectiveness, especially around the sanctions imposed after February two thousand twenty-two.

Catalog record

CREATORArthur Capozzi, Salvatore Vilella, Dario Moncalvo, Marco Fornasiero, Valeria Ricci, Silvia Ronchiadin, Giancarlo Ruffo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-15896

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

WeirdFlows: Anomaly Detection in Financial Transaction Flows Arthur Capozzi, Salvatore Vilella, Dario Moncalvo, Marco Fornasiero, Valeria Ricci, Silvia Ronchiadin, Giancarlo Ruffo Abstract In recent years, the digitization and automation of anti-financial crime (AFC) investigative processes have faced significant challenges, particularly the need for interpretability of AI model results and the lack of labeled data for training. Network analysis has emerged as a valuable approach in this context. In this paper we present WeirdFlows , a top-down search pipeline for detecting potentially fraudulent transactions and non-compliant agents. In a transaction network, fraud attempts are often based on complex transaction patterns that change over time to avoid detection. The WeirdFlows pipeline requires neither an a priori set of patterns nor a training set. In addition, by providing elements to explain the anomalies found, it facilitates and supports the work of an AFC analyst. We evaluate WeirdFlows on a dataset from Intesa Sanpaolo (ISP) bank, comprising 80 million cross-country transactions over 15 months, benchmarking our implementation of the algorithm. The results, corroborated by ISP AFC experts, highlight its effectiveness in identifying suspicious transactions and actors, particularly in the context of the economic sanctions imposed in EU after February 2022. This demonstrates WeirdFlows ’ capability to handle large datasets, detect complex transaction patterns, and provide the necessary interpretability for formal AFC investigations. 1 Introduction and Related Work Financial markets are complex, adaptive systems involving diverse actors such as hedge funds, individual investors, and banks. Complex networks effectively model these systems, focusing on stabilization and destabilization dynamics [ 13 ] , stock price correlations [ 24 ] , shareholder networks [ 4 ] , and the resilience of financial networks [ 11 ] . International trade networks based on linguistic, cultural, or religious ties have also been examined [ 20 ] , along with applications to financial crises, including the 2008 global financial crisis [ 1 ] . Beyond stock markets, network theory models competitive dynamics on the World Wide Web [ 16 ] and the cryptocurrency market, analyzing dominant cryptocurrencies [ 18 ] , transaction structures [ 21 ] , and Ethereum network attributes [ 14 ] . Network analysis is vital for anti-financial crime (AFC) and anti-money laundering (AML), detecting anomalies in transaction networks and automating fraud detection to enhance efficiency. Tools like VISFAN [ 6 ] use network metrics for identifying suspicious transactions. García et al. [ 9 ] applied network analysis in the Spanish Revenue Agency’s Tax Control Study, using algorithms for rapid fraud detection and community detection techniques for economic landscape representation. Colladon et al. [ 7 ] highlighted social network metrics for identifying money laundering through relational graphs of economic sectors, regions, transaction volumes, and ownership links. The CoDetect framework [ 10 ] integrates network and feature data for fraud detection. AMLSim [ 26 ] generates synthetic bank transaction data with identifiable money laundering patterns. Preliminary results indicate that graph learning for AML is feasible even in large, sparse networks, with graph compression techniques like Ligra+ achieving significant compression [ 22 ] . Garcia-Bedoya et al. [ 8 ] advocate for AI and network analysis in AML, identifying three interaction types that conceal money laundering: path (money sent through intermediaries), cycle (money returns to its origin), and smurf (dividing a transaction into smaller ones). Liu et al. [ 15 ] suggest transaction cycles can indicate fraud. Machine learning and deep learning, particularly graph neural networks, are explored for fraud detection [ 5 ] . Unsupervised anomaly detection is often used due to a lack of annotated training data, with innovative approaches like zero-shot learning [ 5 ] . Pan [ 17 ] proposes a deep-set algorithm combining meta-learning and zero-shot learning for money laundering detection. Reviews highlight effective anomaly detection strategies in fraud detection [ 2 , 19 ] , using social network analysis to uncover organized criminal behavior [ 27 ] . 1.1 WeirdFlows In this paper, we introduce WeirdFlows , a top-down search pipeline based on network analysis, designed to aid Anti-Financial Crime (AFC) analysts in detecting illicit transactions and non-compliant agents. We evaluate WeirdFlows on a dataset of 80 million cross-country bank transactions over 15 months, provided anonymously by Intesa Sanpaolo bank, compliant with legal privacy and security standards. This evaluation focuses on identifying attempts to bypass economic sanctions against Russia following the Ukraine invasion on February 24, 2022. Financial fraud often involves complex, evolving transaction patterns to evade detection. WeirdFlows identifies these by defining a transaction flow as a set of payment lines from x to y through intermediaries. Unlike other methods, WeirdFlows does not rely on predefined patterns, addressing the challenge of unlabeled data in AFC tools. Moreover, it provides interpretability, crucial for formal investigations by domain experts, unlike many black-box models [ 5 ] . The application of WeirdFlows involves two steps: 1. Determine Structural Granularity : create a transaction network using financial transactions data. Each node can be any sort of entity, such as an ISO code , BIC , or IBAN . ISO 3166 codes define country names, BIC (Bank Identifier Code) identifies banks in international payments, and IBAN (International Bank Account Number) identifies bank accounts across borders. 2. Compute paths on temporal networks : Generate multiple weighted directed temporal networks by aggregating wire transfers over time (e.g., weekly, monthly). An algorithm computes all possible paths from a selected node for each temporal aggregation. Analysts explore these paths to identify trends, patterns, or anomalies and estimate the total money sent between nodes through intermediaries. Section 3 presents the application of WeirdFlows , demonstrating: 1. Its effectiveness in identifying anomalous transaction flows; 2. Its scalability to millions of transactions; 3. Its interpretability, facilitating formal investigations. We formalize transaction networks and transaction flows in the next section, describe the algorithm for identifying transaction flows, and propose a technique for weighting flows between nodes. Sections 3 covers the application experiments and performance benchmarks of WeirdFlows at different granularities, while the last section discusses the limitation and the possible extensions of the work. 2 Methodology 2.1 Weighing a Transaction Flow Transactions networks, usually represented as weighted directed temporal networks, are very useful to model financial transactions over time. On such a network G_{T}=(V,L_{T}) , we can define a path as an ordered finite collection of n distinct edges connecting vertices i and j during interval T : \begin{split}P^{T}_{i,j}=\{&e(i,v_{1}),e(v_{1},v_{2}),e(v_{2},v_{3}),\ldots,e(v_{n-1},j)\},\\ &\text{with }v_{x}\in V\text{ and }e(v_{x},v_{x+1})\in L_{T}\\ &\text{for all }x\in\{1,\ldots,n-1\}\end{split} (1) The weight of the path P^{T}_{i,j} is then given by \sum_{i=1}^{n-1}w(e(v_{i},v_{i+1})) (2) and W(P_{i})=\{w(e(v_{1},v_{2})),...,w(e(v_{n-1},v_{n}))\} is the set containing the weights of the edges of the path P_{i} . Finally, \text{Paths}^{n}_{i,j}(G_{T}) is the set of all possible paths of length n from i to j in graph G at interval T . In real-world scenarios, to avoid detection of fraudulent activity, an agent may use intermediaries to transfer money. The number of intermediary groups or the length of each path connecting any two nodes x to y often cannot be known a priori. A transaction through n intermediaries in a transaction network forms a path of length n+1 . The WeirdFlows pipeline aims to identify the transaction flow from a given input node x and verify the maximum possible amount of money sent from x to other nodes

Read the complete text