Tide: A Customisable Dataset Generator for Anti-Money Laundering Research

Montijn van den Beukel; Jože Martin Rožanec; Ana-Lucia Varbanescu

cs.LGMar 2, 2026

Tide: A Customisable Dataset Generator for Anti-Money Laundering Research

Authors: Montijn van den Beukel, Jože Martin Rožanec, Ana-Lucia Varbanescu

Organizations: University of Amsterdam, The Netherlands · Jožef Stefan Institute, Slovenia · University of Amsterdam/University of Twente, The Netherlands

Abstract

The lack of accessible transactional data significantly hinders machine learning research for Anti-Money Laundering (AML). Privacy and legal concerns prevent the sharing of real financial data, while existing synthetic generators focus on simplistic structural patterns and neglect the temporal dynamics (timing and frequency) that characterise sophisticated laundering schemes. We present Tide, an open-source synthetic dataset generator that produces graph-based financial networks incorporating money laundering patterns defined by both structural and temporal characteristics. Tide enables reproducible, customisable dataset generation tailored to specific research needs. We release two reference datasets with varying illicit ratios (LI: 0.10%, HI: 0.19%), alongside the implementation of state-of-the-art detection models. Evaluation across these datasets reveals condition-dependent model rankings: LightGBM achieves the highest PR-AUC (78.05) in the low illicit ratio condition, while XGBoost performs best (85.12) at higher fraud prevalence. These divergent rankings demonstrate that the reference datasets can meaningfully differentiate model capabilities across operational conditions. Tide provides the research community with a configurable benchmark that exposes meaningful performance variation across model architectures, advancing the development of robust AML detection methods.

Figures & tables

Feature	AMLSim ( Suzumura and Kanezashi, 2021 )	AMLWorld ( Altman and others, 2023 )	SynthAML ( Jensen et al., 2023 )	Tide (Ours)
Core Approach	Agent-Based	Virtual World Agents	Statistical (Copulas)	Pattern Injection
Executable Generator	✓	✗	✗	✓
Pre-loaded Pattern Library	✓ 1	✓ 1	✗	✓ 2
Custom Pattern Injection	✗	✗	✗	✓
Precise Temporal Control	✗	✗	✗	✓
Zero-Shot (No Real Data)	✓	✓	✗	✓

Table 1. Comparison of synthetic financial transaction generators. Unlike predecessors that offer either fixed patterns (AMLSim) or statistical approximations (SynthAML), Tide provides a dual capability: it is preloaded with a library of complex, well-researched typologies (e.g., U-Turns, Smurfing) while simultaneously allowing the programmatic injection of novel, user-defined patterns.

Panel of Figure 1 — Figure 1. The iterative research methodology. The process employs an adversarial feedback loop where generator parameters (Phase 2) are recalibrated based on the performance of detection models (Phase 4).

Attribute	Description	Type
node_id	Unique identifier for each node.	String
node_type	The type of the entity.	Categorical
country_code	Country where the entity is based.	String
is_fraudulent	Label indicating whether node is part of a fraud pattern.	Boolean
risk_score	Value in [0, 0.9] expressing the probability of an entity being fraudulent.	Float

Table 2. Common attributes for all node types.

Component	Description	Type
$i,j$	Source and target node IDs.	String
$r$	Relationship type: transaction or ownership .	Categorical
$a$	Transaction amount.	Float
$t$	Timestamp of the interaction.	Timestamp
$\delta$	Time since the source node’s last transaction.	Timedelta
$\gamma$	Transaction category.	Categorical

Table 3. Attributes of an edge tuple

e=(i,j,r,a,t,\delta,\gamma,\lambda)

Panel of Figure 2 — Figure 2. The four main steps of the Tide graph generation process: (a) Entity creation and clustering, (b) Entity selection, (c) Transaction sequence generation, and (d) Pattern aggregation.

Figure 3. Visualisation of the overseas transfer pattern. Individual_1 owns Account_1 .

Figure 4. Visualisation of the rapid fund movement pattern. Individual_1 owns Account_1 and Account_2 .

Figure 5. Visualisation of the front business pattern. Business_1 owns Account_1 and Account_2 .

Figure 6. Visualisation of the U-turn transaction pattern.

Type	$\mu$	$\sigma$	Source Avg	Gen. Median	Gen. Mean
Payment	3.8	1.2	$60 (FRPS)	$46	$218
Transfer	5.5	2.0	$1,634 (FRPS)	$221	$1,863
Withdrawal	4.8	0.9	$198 (FRPS)	$130	$546
Deposit †	5.3	1.5	—	$220	$977

Table 4. Log-normal amount parameters and generated statistics per transaction type in LI dataset.

Type	FRPS Share †	Generated Share
Payment	$\sim$ 91%	68%
Transfer	$\sim$ 7%	12%
Withdrawal	$\sim$ 2%	8%
Deposit	—	11%

Table 5. Transaction type frequencies: FRPS reference vs. generated data in LI dataset.

Variant	Nodes	Edges	Illicit ratio	Imbalance
LI	36,629	7,642,030	0.10%	959:1
HI	36,653	7,618,299	0.19%	531:1

Table 6. Published dataset variants. Both share the same base graph (8,000 individuals, 12-month simulation).

Panel of Figure 7 — Figure 7. The front business pattern shown as a network graph and a transaction timeline.

Panel of Figure 8 — Figure 8. The rapid fund movement pattern shown as a network graph and a transaction timeline.

Panel of Figure 9 — Figure 9. The repeated overseas transfer pattern shown as a network graph and two distinct transaction timelines: one for high-frequency activity and one for periodic behaviour.

Panel of Figure 10 — Figure 10. The U-turn transaction pattern shown as a network graph and a transaction timeline.

Model	Illicit Ratio
	LI (0.10%)	HI (0.19%)
GIN	$69.86\pm 6.55$	$71.14\pm 8.70$
GIN+EU	$39.43\pm 20.99$	$48.42\pm 17.05$
PNA	$69.20\pm 0.00$	$75.63\pm 5.85$
LightGBM+GFP	$21.45\pm 13.23$	$64.40\pm 24.36$
XGBoost+GFP	$35.45\pm 2.59$	$32.20\pm 1.68$

Table 7. PR-AUC across experimental conditions. Bold indicates best performance per condition. Results are mean

\pm

standard deviation over 5 runs.

Attribute	Description	Type
Name	An individual’s full name.	String
Age group	The individual’s age group.	Categorical
Occupation	The individual’s occupation.	Categorical
Gender	The individual’s gender.	Categorical

Table 8. Specific attributes for nodes with the individual entity type.

Attribute	Description	Type
business_category	The primary sector or type of business activity.	Categorical
incorporation_year	The year the business was legally incorporated.	Integer
number_of_employees	Approximate number of people employed by the business.	Integer
is_high_risk_category	Flag for businesses in a high-risk category.	Boolean

Table 9. Specific attributes for nodes with the business entity type.

Attribute	Description	Type
account_category	Type of account (e.g., Current, Savings, Cash).	Categorical
currency	The currency used by the account.	String
owner_id	ID of the Individual or Business that owns this account.	String
institution_id	ID of the Financial Institution that hosts this account.	String

Table 10. Specific attributes for nodes with the account entity type.

Attribute	Description	Type
institution_name	Name of the financial institution.	String
country	Country of legal registration or regulation.	String

Table 11. Specific attributes for nodes with the financial institution entity type.

Entity Type	Risk Factor	Weight ( $w_{i}$ )
Individual	Base risk	0.05
	High-risk age group	0.15
	High-risk occupation	0.12
Business	Base risk	0.10
	Cash-intensive business category	0.25
	Very small company ( $\leq$ 5 employees)	0.10

Table 12. Default risk factor weights used in entity risk scoring.

Pattern	Topology	Timing	Additional layering
Overseas Transfers	1 → n overseas (n=2–5)	Periodic (7/14/30d)	✓
Rapid Movement	n → 1 → cash-out (n=2–7)	Burst (1–24h delay)	✓
Front Business	Cash → business → overseas	Paired (0.5–6h)	✓
U-Turn	Source → chain → return	Sequential (1–5d/hop)	Inherent
Synchronised	n → 1 recipient (n=3–8)	Coordinated (2h window)	—

Table 13. Overview of implemented laundering patterns.

Pattern	Group	Fraud feature countered
Random everyday payments	Baseline	—
High-value transactions	Baseline	—
Salary payments	Baseline	—
Fraudster background	Baseline	Behavioural anomaly
Legitimate bursts	Counter-leakage	Temporal clustering
Legitimate periodic	Counter-leakage	Periodic timing

Table 14. Overview of background activity patterns. Baseline patterns model routine financial behaviour; counter-leakage patterns introduce legitimate activity that shares features with fraud patterns.

Table 15. Node and edge features included in the exported dataset. Categorical features are one-hot encoded. Internal generation attributes ( risk_score , occupation , is_high_risk_category , etc.) are excluded.

Parameter	Value
Batch size	128
Patterns	Scatter-gather, temporal cycles,
	simple cycles, vertex statistics
Scatter-gather window	6 hours
Cycle / vertex stats window	1 day
Max simple cycle length	10

Table 16. Graph Feature Preprocessor (GFP) configuration, adopted from Altman et al. ( Altman and others, 2023 ) .

FRPS Category	Volume (B)	Share
Non-prepaid debit cards	87.8	42.2%
Credit cards	51.1	24.6%
ACH debit transfers	20.3	9.8%
Prepaid debit cards	18.1	8.7%
ACH credit transfers	15.9	7.6%
Checks	11.2	5.4%

Table 17. FRPS 2021 noncash payment volumes by category.

Parameter	LightGBM	XGBoost
n_estimators	1000	1000
max_depth	6	6
num_leaves	64	–
learning_rate	0.01	0.01
reg_lambda (L2)	5.0	5.0
reg_alpha (L1)	1.5	0.5

Table 18. Hyperparameter configuration for GBT models.

Parameter	GIN	GIN+EU	PNA
hidden dimension	64	64	64
number of layers	4	4	2
dropout	0.2	0.2	0.1
final dropout	0.5	0.5	0.5
learning rate	0.001
weight decay	0.0005

Table 19. Hyperparameter configuration for GNN models.

Model	F1	Precision	Recall	PR-AUC	Lift	Youden’s J	ROC-AUC
GIN	$5.95\pm 1.27$	$3.07\pm 0.67$	$94.33\pm 2.62$	$69.86\pm 6.55$	$670.63\pm 62.91$	$90.66\pm 3.38$	$97.64\pm 1.05$
GIN+EU	$4.45\pm 3.09$	$2.31\pm 1.66$	$87.55\pm 7.82$	$39.43\pm 20.99$	$378.46\pm 201.54$	$80.91\pm 10.56$	$94.66\pm 3.99$
PNA	$69.90\pm 0.00$	$57.86\pm 0.00$	$88.26\pm 0.00$	$69.20\pm 0.00$	$664.30\pm 0.00$	$88.19\pm 0.00$	$98.44\pm 0.00$
LightGBM+GFP	$12.13\pm 7.18$	$6.64\pm 4.20$	$92.81\pm 8.64$	$21.45\pm 13.23$	$205.95\pm 127.04$	$90.12\pm 10.92$	$96.60\pm 5.38$
XGBoost+GFP	$10.19\pm 2.04$	$5.38\pm 1.15$	$99.31\pm 0.40$	$35.45\pm 2.59$	$340.28\pm 24.88$	$97.17\pm 0.20$	$99.77\pm 0.02$
LightGBM	$8.01\pm 2.21$	$4.20\pm 1.20$	$95.73\pm 3.50$	$78.05\pm 3.61$	$749.19\pm 34.64$	$92.80\pm 3.16$	$98.80\pm 1.75$

Table 20. Comparison of model performance on fraud detection. Homophily index: 0.0023. Class imbalance of 0.10%. Lift = PR-AUC / fraud_rate (times better than random). All metrics (F1, Precision, Recall) computed at Youden’s optimal threshold . Results reported as mean

\pm

std over 5 runs.

Model	F1	Precision	Recall	PR-AUC	Lift	Youden’s J	ROC-AUC
GIN	$9.83\pm 2.35$	$5.20\pm 1.29$	$91.65\pm 5.28$	$71.14\pm 8.70$	$378.40\pm 46.29$	$87.98\pm 6.14$	$96.71\pm 1.90$
GIN+EU	$4.88\pm 2.05$	$2.52\pm 1.09$	$86.59\pm 5.20$	$48.42\pm 17.05$	$257.56\pm 90.71$	$78.49\pm 7.27$	$93.61\pm 2.85$
PNA	$31.67\pm 13.94$	$20.20\pm 10.83$	$91.42\pm 1.34$	$75.63\pm 5.85$	$402.30\pm 31.14$	$90.31\pm 0.57$	$98.61\pm 0.22$
LightGBM+GFP	$20.98\pm 10.06$	$12.13\pm 6.46$	$97.35\pm 1.84$	$64.40\pm 24.36$	$342.56\pm 129.57$	$95.28\pm 2.21$	$99.11\pm 0.94$
XGBoost+GFP	$12.46\pm 0.65$	$6.65\pm 0.37$	$97.90\pm 0.28$	$32.20\pm 1.68$	$171.26\pm 8.93$	$95.06\pm 0.12$	$99.53\pm 0.03$
LightGBM	$28.59\pm 8.68$	$17.04\pm 5.72$	$96.69\pm 2.31$	$70.47\pm 27.91$	$374.85\pm 148.46$	$95.51\pm 2.97$	$99.45\pm 0.68$

Table 21. Comparison of model performance on fraud detection. Homophily index: 0.0035. Class imbalance of 0,19%. Lift = PR-AUC / fraud_rate (times better than random). All metrics (F1, Precision, Recall) computed at Youden’s optimal threshold . Results reported as mean

\pm

std over 5 runs.

Table 22. Performance comparison of PNA and XGBoost models on the classification task on the HI dataset. Precision-Recall curves show Average Precision (AP) and F1 scores at optimal thresholds.

Configuration	Transactions	Total ( $N$ )	Wall-time	Peak Memory
	[millions]	[millions]	[s]	[MB]
Test-1k	$0.372\pm 0.004$	$0.377\pm 0.004$	$2.3\pm 0.3$	$490\pm 30$
Test-2k	$0.948\pm 0.005$	$0.957\pm 0.005$	$5.2\pm 0.3$	$1090\pm 30$
Test-3.5k	$1.99\pm 0.01$	$2.00\pm 0.01$	$11.0\pm 0.9$	$2200\pm 50$
Test-5k	$3.28\pm 0.03$	$3.30\pm 0.03$	$18\pm 1$	$2700\pm 200$
Test-8k	$5.99\pm 0.02$	$6.03\pm 0.02$	$43\pm 3$	$3600\pm 40$

Table 23. Aggregated generation performance across different graph scales on an M2 MacBook Air with 8 GB of RAM. Results are presented as mean

\pm

standard deviation over 3 runs.

Panel of Figure 11 — Figure 11. Performance analysis of the graph generator. (a) Log-log plot showing generation time versus total graph elements with near-linear scaling (exponent $\alpha=1.164$ , $R^{2}=0.976$ ). (b) Average throughput in kElements/s. (c) Peak memory usage scaling with sub-linear exponent of 0.627. (d) Memory efficiency showing additional memory required per element. Gray dotted lines in (a) and (c) represent $\mathcal{O}(N)$ for reference.

Metric	Value	Target
Pearson correlation ( $r$ )	0.970	$>0.9$
$p$ -value	0.00137	$<0.05$
Scaling exponent ( $\alpha$ )	1.164	$\alpha\leq 1.5$
Coefficient of determination ( $R^{2}$ )	0.976	$>0.9$

Table 24. Statistical metrics from scalability analysis.

Anti-Money Laundering Detection

See all 20 papers

Apr 19, 2026

TransXion: A High-Fidelity Graph Benchmark for Realistic Anti-Money Laundering

Money laundering poses severe risks to global financial systems, driving the widespread adoption of machine learning for transaction monitoring. However, progress remains stifled by the lack of realistic benchmarks. Existing transaction-graph datasets suffer from two pervasive limitations: (i) they provide sparse node-level semantics beyond anonymized identifiers, and (ii) they rely on template-driven anomaly injection, which biases benchmarks toward static structural motifs and yields overly optimistic assessments of model robustness. We propose TransXion, a benchmark ecosystem for Anti-Money Laundering (AML) research that integrates profile-aware simulation of normal activity with stochastic, non-template synthesis of illicit subgraphs.TransXion jointly models persistent entity profiles and conditional transaction behavior, enabling evaluation of "out-of-character" anomalies where observed activity contradicts an entity's socio-economic context. The resulting dataset comprises approximately 3 million transactions among 50,000 entities, each endowed with rich demographic and behavioral attributes. Empirical analyses show that TransXion reproduces key structural properties of payment networks, including heavy-tailed activity distributions and localized subgraph structure. Across a diverse array of detection models spanning multiple algorithmic paradigms, TransXion yields substantially lower detection performance than widely used benchmarks, demonstrating increased difficulty and realism. TransXion provides a more faithful testbed for developing context-aware and robust AML detection methods. The dataset and code are publicly available at https://github.com/chaos-max/TransXion.

Keyang Chen, Mingxuan Jiang, Yongsheng Zhao +9

Fudan University · Peking University

Financial Fraud Detection Synthetic Benchmark Generation Anti-Money Laundering Synthetic Anomaly Generation Anomaly Detection Anti-Money Laundering Detection

Ask PDF

Apr 3, 2026

Extracting Money Laundering Transactions from Quasi-Temporal Graph Representation

Money laundering presents a persistent challenge for financial institutions worldwide, while criminal organizations constantly evolve their tactics to bypass detection systems. Traditional anti-money laundering approaches mainly rely on predefined risk-based rules, leading to resource-intensive investigations and high numbers of false positive alerts. In order to restrict operational costs from exploding, while billions of transactions are being processed every day, financial institutions are investing in more sophisticated mechanisms to improve existing systems. In this paper, we present ExSTraQt (EXtract Suspicious TRAnsactions from Quasi-Temporal graph representation), an advanced supervised learning approach to detect money laundering (or suspicious) transactions in financial datasets. Our proposed framework excels in performance, when compared to the state-of-the-art AML (Anti Money Laundering) detection models. The key strengths of our framework are sheer simplicity, in terms of design and number of parameters; and scalability, in terms of the computing and memory requirements. We evaluated our framework on transaction-level detection accuracy using a real dataset; and a set of synthetic financial transaction datasets. We consistently achieve an uplift in the F1 score for most datasets, up to 1% for the real dataset; and more than 8% for one of the synthetic datasets. We also claim that our framework could seamlessly complement existing AML detection systems in banks. Our code and datasets are available at https://github.com/mhaseebtariq/exstraqt.

Haseeb Tariq, Marwan Hassani

ING Bank, Amsterdam, The Netherlands · Eindhoven University of Technology, Eindhoven, The Netherlands

Anti-Money Laundering Detection Anti-Money Laundering Financial Fraud Detection

Ask PDF

Apr 1, 2026

Detecting Complex Money Laundering Patterns with Incremental and Distributed Graph Modeling

Money launderers take advantage of limitations in existing detection approaches by hiding their financial footprints in a deceitful manner. They manage this by replicating transaction patterns that the monitoring systems cannot easily distinguish. As a result, criminally gained assets are pushed into legitimate financial channels without drawing attention. Algorithms developed to monitor money flows often struggle with scale and complexity. The difficulty of identifying such activities is further intensified by the (persistent) inability of current solutions to control the excessive number of false positive signals produced by rigid, risk-based rules systems. We propose a framework called ReDiRect (REduce, DIstribute, and RECTify), specifically designed to overcome these challenges. The primary contribution of our work is a novel framing of this problem in an unsupervised setting; where a large transaction graph is fuzzily partitioned into smaller, manageable components to enable fast processing in a distributed manner. In addition, we define a refined evaluation metric that better captures the effectiveness of exposed money laundering patterns. Through comprehensive experimentation, we demonstrate that our framework achieves superior performance compared to existing and state-of-the-art techniques, particularly in terms of efficiency and real-world applicability. For validation, we used the real (open source) Libra dataset and the recently released synthetic datasets by IBM Watson. Our code and datasets are available at https://github.com/mhaseebtariq/redirect.

Haseeb Tariq, Alen Kaja, Marwan Hassani

ING Bank Amsterdam, The Netherlands · Eindhoven University of Technology Eindhoven, The Netherlands · Identity & Risk Intelligence Adyen Amsterdam, The Netherlands +1

Anti-Money Laundering Detection Anti-Money Laundering Financial Fraud Detection Anomaly Detection Unsupervised Anomaly Detection

Ask PDF

Abstract

Anti-Money Laundering Detection

Explore similar work

TransXion: A High-Fidelity Graph Benchmark for Realistic Anti-Money Laundering

Extracting Money Laundering Transactions from Quasi-Temporal Graph Representation

Detecting Complex Money Laundering Patterns with Incremental and Distributed Graph Modeling