Paper: arXiv 2306.16424

Authors: Erik Altman, Jovan Blanuša, Luc von Niederhäusern, Béni Egressy, Andreea Anghel, Kubilay Atasu

Abstract

With the widespread digitization of finance and the increasing popularity of cryptocurrencies, the sophistication of fraud schemes devised by cybercriminals is growing. Money laundering – the movement of illicit funds to conceal their origins – can cross bank and national boundaries, producing complex transaction patterns. The UN estimates 2-5% of global GDP or $0.8 - $2.0 trillion dollars are laundered globally each year. Unfortunately, real data to train machine learning models to detect laundering is generally not available, and previous synthetic data generators have had significant shortcomings. A realistic, standardized, publicly-available benchmark is needed for comparing models and for the advancement of the area. To this end, this paper contributes a synthetic financial transaction dataset generator and a set of synthetically generated AML (Anti-Money Laundering) datasets. We have calibrated this agent-based generator to match real transactions as closely as possible and made the datasets public. We describe the generator in detail and demonstrate how the datasets generated can help compare different machine learning models in terms of their AML abilities. In a key way, using synthetic data in these comparisons can be even better than using real data: the ground truth labels are complete, whilst many laundering transactions in real data are never detected.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper introduces a complex agent-based simulation to generate realistic synthetic datasets and benchmarks with graph neural networks, but its primary focus is on practical dataset creation and benchmarking rather than theoretical derivations.

Research Flowchart

  flowchart TD
  A["Research Goal: Create a realistic<br>public benchmark for AML models"] --> B["Agent-Based Data Generation"]
  
  B --> C["Calibrated to match<br>real transaction patterns"]
  C --> D["Synthetic Datasets<br>with complete ground truth"]
  
  D --> E["Computational Process:<br>ML Model Comparison"]
  E --> F["Key Findings:<br>1. Public AML benchmark dataset created<br>2. Superior to real data for evaluation<br>3. Ground truth is complete vs real data<br>4. Enables model advancement"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a synthetic financial transaction generator and a set of realistic, standardized AML datasets for benchmarking anti‑money‑laundering models.

Source sentence
A synthetic financial transaction generator, which builds a multi-agent virtual world in which some of the agents are criminals with illicit income to launder

Assumes 5

  • The synthetic generator is calibrated to match real transactions as closely as possible. model
    Source sentence
    We have calibrated this agent-based generator to match real transactions as closely as possible and made the datasets public.
  • The generator knows the source of all funds, providing perfect labeling of laundering transactions. data
    Source sentence
    Since the generator knows the source of all funds, it knows which transfers are laundering.
  • The generated AML datasets are highly imbalanced between laundering and non‑laundering transactions. data
    Source sentence
    As our proposed datasets are naturally very imbalanced, traditional metrics like accuracy do not provide a reliable measure of model performance.
  • Criminal agents launder less frequently in low‑illicit (LI) datasets, making pattern discovery harder. model
    Source sentence
    The primary difference between LI and HI datasets is that criminals launder less frequently in LI, which makes it harder to discover the laundering patterns.
Show 1 more
  • Weighting the minority class higher in the loss function improves performance on LI datasets. model
    Source sentence
    To address this, our models incorporate the measure of weighting the predictions of the minority class higher in the loss function.

Data

  • AMLworld synthetic AML datasets · Kaggle synthetic public
    Source sentence
    Six AMLworld datasets are now publicly available on Kaggle [50].

Recognised sources: Kaggle

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Message‑passing GNNs that exploit account connectivity capture laundering transactions effectively, with advanced architectures (PNA, GIN+EU) significantly enhancing performance. performance
    Source sentence
    These results show that message-passing GNNs that exploit the connectivity between accounts are effective in capturing laundering transactions. Advanced GNN architectures such as PNA and GIN+EU, which respectively implement extended message aggregation and edge update mechanisms, significantly enhance the GNN performan
  • Pre‑training on higher‑illicit (HI) datasets and fine‑tuning improves F1 on lower‑illicit (LI) datasets, e.g., PNA fine‑tuned reaches 27.38 % F1 on LI‑Small. performance
    Source sentence
    ... performance on LI data can be improved further for the LI‑Small dataset, where directly using the models trained on HI datasets is ineffective, by fine-tuning the HI PNA models using the LI dataset.
  • GBT baselines achieve competitive results, with near‑par performance to PNA despite lacking handcrafted features. performance
    Source sentence
    An interesting observation is the near-par performance of PNA with GBT baselines, even without relying on handcrafted features.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 13 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive (2 by independent authors): Graph Neural Networks for Financial Fraud Detection: A … (uses its method or data)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.