README.md

metadata

language:
  - en
license: cc-by-4.0
task_categories:
  - tabular-classification
  - time-series-forecasting
size_categories:
  - 1M<n<10M
tags:
  - fraud-detection
  - payments
  - synthetic
  - SEPA
  - anomaly-detection
  - finance
  - tabular
pretty_name: SynSEPA — Synthetic SEPA Instant Payment Fraud Dataset
configs:
  - config_name: transactions
    default: true
    data_files:
      - split: train
        path: data/synsep_full_dataset.csv
  - config_name: accounts
    data_files:
      - split: train
        path: data/accounts.csv
dataset_info:
  - config_name: transactions
    description: >
      SynSEPA is a synthetic SEPA Instant Credit Transfer dataset containing
      1.84M transactions across 10,000 accounts, with 4 APP fraud typologies
      injected following EPC 2025 definitions. Calibrated against EBA-ECB 2025
      fraud statistics.
    features:
      - name: transaction_id
        dtype: string
      - name: account_id
        dtype: string
      - name: persona
        dtype: string
      - name: timestamp
        dtype: string
      - name: sender_iban
        dtype: string
      - name: beneficiary_iban
        dtype: string
      - name: beneficiary_country
        dtype: string
      - name: country_type
        dtype: string
      - name: amount
        dtype: float64
      - name: remittance_category
        dtype: string
      - name: remittance_text
        dtype: string
      - name: hour_of_day
        dtype: int64
      - name: day_of_week
        dtype: int64
      - name: is_weekend
        dtype: int64
      - name: time_since_last_txn
        dtype: float64
      - name: is_new_beneficiary
        dtype: int64
      - name: is_fraud
        dtype: int64
      - name: fraud_type
        dtype: string
    splits:
      - name: train
        num_examples: 1839560
  - config_name: accounts
    description: >
      Per-account metadata for the 10,000 synthetic SynSEPA accounts — persona,
      home country, behavioural parameters, and known-beneficiary lists.
    features:
      - name: account_id
        dtype: string
      - name: persona
        dtype: string
      - name: description
        dtype: string
      - name: home_country
        dtype: string
      - name: sender_iban
        dtype: string
      - name: txn_per_month_min
        dtype: int64
      - name: txn_per_month_max
        dtype: int64
      - name: typical_amount
        dtype: float64
      - name: foreign_txn_prob
        dtype: float64
      - name: weekend_factor
        dtype: float64
      - name: fraud_target_types
        dtype: string
      - name: known_beneficiaries
        dtype: string
      - name: known_beneficiary_count
        dtype: int64
    splits:
      - name: train
        num_examples: 10000

SynSEPA: A Synthetic SEPA Instant Payment Dataset for APP Fraud Detection Research

---

Overview

SynSEPA is the first publicly available synthetic dataset specifically designed for Authorised Push Payment (APP) fraud detection in SEPA Instant Credit Transfer payments.

No public SEPA fraud dataset previously existed. SynSEPA fills this gap by providing a large-scale, statistically calibrated synthetic dataset grounded in official EU regulatory statistics from the EBA-ECB 2025 Payment Fraud Report and the EPC 2025 Payment Threats and Fraud Trends Report.

The dataset is released as part of the SEPAGen research project: "Transformer-Based Generative Anomaly Detection for APP Fraud in SEPA Instant Payments"

---

Why SynSEPA?

The Problem

• SEPA Instant payments settle in under 10 seconds and are irreversible
• APP fraud — where victims are manipulated into authorising payments — caused €2.5 billion in EU losses in 2024 (EBA-ECB 2025), up 24% YoY
• No public SEPA fraud dataset existed for training or benchmarking ML models
• Real SEPA fraud data is proprietary, privacy-sensitive, and inaccessible to researchers

The Solution

SynSEPA generates realistic SEPA payment sequences with:

• 4 APP fraud typologies as defined by the EPC 2025 report
• 4 customer personas representing the EU retail banking population
• Statistical calibration against EBA-ECB 2025 fraud report benchmarks
• Sequential structure enabling sequence-based generative model training

---

Dataset Statistics

Property	Value
Total transactions	1,839,560
Normal transactions	1,832,448 (99.61%)
Fraud transactions	7,112 (0.39%)
Unique accounts	10,000
Simulation period	January – December 2024
Countries covered	10 EU/EEA countries
Avg transactions per account	184
Min transactions per account	69

Fraud Breakdown

Typology	Transactions	Victims	Avg Txns/Victim
Romance Scam	4,232	720	5.9
Bank Impersonation	1,440	1,440	1.0
Invoice / Mandate Fraud	900	900	1.0
CEO / BEC Fraud	540	540	1.0
Total	7,112	3,600	1.98

Account Personas

Persona	Accounts	%	Fraud Targets
Regular Employee	5,054	50.5%	Impersonation, Romance
Student	1,972	19.7%	Romance, Impersonation
Retiree	1,519	15.2%	Impersonation, Romance
Small Business	1,455	14.5%	Invoice, CEO

---

Schema — synsep_full_dataset.csv

Core SEPA Fields

Column	Type	Description	Example
transaction_id	string	Unique transaction ID	TXN_00000001
account_id	string	Sender account ID	ACC_000042
persona	string	Account type	employee
timestamp	datetime	Transaction datetime	2024-03-15 14:23:01
sender_iban	string	Sender IBAN (synthetic)	DE89370400440532013000
beneficiary_iban	string	Beneficiary IBAN (synthetic)	FR7630006000011234
beneficiary_country	string	Beneficiary country code	FR
amount	float	Transaction amount (EUR)	245.50
remittance_category	string	Payment category	grocery
remittance_text	string	Payment reference text	Rent Mar 2024

Engineered Features

Column	Type	Description	Example
country_type	string	domestic / eu_cross_border / non_eu	domestic
hour_of_day	int	Hour of transaction (0-23)	14
day_of_week	int	Day of week (0=Monday, 6=Sunday)	2
is_weekend	int	Weekend flag (0/1)	0
time_since_last_txn	float	Seconds since previous transaction	86400.0
is_new_beneficiary	int	New IBAN never seen before (0/1)	0

Labels (for evaluation only — never used in unsupervised training)

Column	Type	Description	Values
is_fraud	int	Fraud label	0 = normal, 1 = fraud
fraud_type	string	Fraud typology	none / impersonation / invoice / romance / ceo

---

Schema — accounts.csv

Column	Type	Description
account_id	string	Unique account ID
persona	string	Account type (employee/student/retiree/business)
home_country	string	Sender's home country
sender_iban	string	Account IBAN
txn_per_month_min/max	int	Transaction frequency range
typical_amount	float	Typical transaction amount
foreign_txn_prob	float	Cross-border transaction probability
fraud_target_types	JSON	Which fraud typologies target this account
known_beneficiaries	JSON	List of regular beneficiary IBANs

---

APP Fraud Typologies in SynSEPA

Why these four typologies?

The four fraud types in SynSEPA were selected directly from the EPC 2025 Payment Threats and Fraud Trends Report (EPC162-24 v2.0), which is the authoritative classification of APP fraud in the SEPA ecosystem published by the European Payments Council. These are not hypothetical categories — they represent the dominant, documented patterns of Authorised Push Payment fraud occurring across EU/EEA markets today.

The selection criteria were:

1. Coverage of the victim population — the four typologies together target all four account personas in SynSEPA (retail employees, students, retirees, and small businesses), ensuring the dataset is not biased toward a single demographic.

2. Diversity of behavioural signatures — each typology produces a distinct statistical footprint in the transaction data (amount level, timing, beneficiary geography, sequence length), making the dataset useful for evaluating whether a model can distinguish between fundamentally different fraud mechanisms rather than just detecting one kind of outlier.

3. Relevance to SEPA Instant specifically — typologies were chosen where the irrevocability and speed of SEPA Instant payments are part of the fraud mechanism itself (e.g. impersonation fraudsters urgently pressure victims to transfer before the bank can intervene; romance scammers exploit the ease of cross-border SEPA transfers). Typologies that primarily exploit card networks or slower payment rails were excluded.

4. Volume calibration against EBA-ECB 2025 — victim counts and fraud volumes per typology were set to reflect the proportions reported in the EBA-ECB 2025 Joint Report on Payment Fraud, so the dataset mirrors the real EU fraud landscape.

Typology Details

1. Bank / Authority Impersonation

Fraudster poses as the victim's bank fraud team or a law enforcement agency, creating urgency around a supposed security threat. Victim is convinced to transfer funds to a "safe account" controlled by the fraudster.

• Target personas: Employees, Retirees, Students
• Behavioural signature: Single large transaction (3–10× the victim's normal amount), new domestic IBAN never seen in history, urgent remittance text ("safe account transfer", "security hold"), sent shortly after the last normal transaction
• Why hard to detect: Amount is large but the IBAN is domestic, and the timing follows a normal inter-transaction gap — it does not look like an unusual payment channel

2. Invoice / Mandate Fraud

Fraudster intercepts a legitimate supplier invoice (via email compromise or postal interception) and replaces the beneficiary IBAN with one they control. The victim pays what they believe is a routine business invoice.

• Target personas: Small Business only
• Behavioural signature: Amount closely mirrors the victim's typical supplier payment (deliberately subtle), new IBAN despite a familiar-looking remittance reference, business accounts only
• Why hard to detect: This is the most subtle typology — the amount is not anomalous, the remittance text looks normal, and only the IBAN is new. A purely amount-based detector will miss it entirely.

3. Romance Scam

Fraudster cultivates a fake online relationship over weeks or months, then gradually requests money under emotional pretexts (medical emergency, travel costs, investment opportunity).

• Target personas: Employees, Students, Retirees
• Behavioural signature: 4–8 escalating payments to the same foreign IBAN over a 5–12 week period, emotional or personal remittance text, cross-border destination
• Why hard to detect: Each individual payment may not look anomalous in isolation — it is the multi-week sequence of escalating payments to the same new foreign IBAN that constitutes the fraud signal. This typology tests whether a model can detect account-level behavioural drift over time.

4. CEO / Business Email Compromise (BEC)

Fraudster impersonates a company's CEO or senior executive via email and instructs a finance employee to make an urgent, confidential international transfer.

• Target personas: Small Business only
• Behavioural signature: Large amount, new non-EU IBAN, Friday afternoon timing (when management is less available to verify), confidential remittance text
• Why hard to detect: The payment instruction appears to come from internal authority. The fraud signal is in the combination of a large non-EU transfer on a Friday with a new IBAN — no single feature is sufficient.

---

Generation Methodology

SynSEPA was generated using a 3-step pipeline (code in code/generate/):

Step 1 — Account Generation

10,000 synthetic accounts assigned to one of 4 personas. Each account receives:

• A syntactically correct IBAN for their home country
• Behavioural parameters (frequency, amount ranges, active hours)
• A list of 5-20 known beneficiary IBANs

Step 2 — Normal Transaction Generation

12 months (Jan-Dec 2024) of transaction history per account, with:

• Transaction frequency sampled from persona distribution
• Timestamps weighted by persona's active hours and day-of-week patterns
• Beneficiaries drawn 80% from known list, 20% new
• Remittance categories and text sampled from persona's profile
• All engineered features computed (time_since_last_txn, is_new_beneficiary, etc.)

Step 3 — Fraud Injection

APP fraud transactions injected following EPC 2025 typology signatures:

• Victim accounts selected by eligible persona type
• Fraud transaction parameters derived from victim's normal history
• Romance scam sequences span 4-8 payments over 5-12 week periods

---

Statistical Calibration

SynSEPA is calibrated against the following EBA-ECB 2025 benchmarks:

Metric	EBA Benchmark	SynSEPA	Status
Fraud rate (volume)	~0.200%	0.387%	Within range
Cross-border rate (normal)	~11%	11.1%	Matches
Weekend transaction rate	<25%	19.8%	Passes
Business hours concentration	>60%	89.1%	Passes
Fraud cross-border rate	> normal	96.4% vs 11.1%	Correct

Note: Fraud rate (0.387%) exceeds EBA's 0.200% volume benchmark primarily because romance scams generate multiple transactions per victim (avg 5.9). This reflects the multi-event nature of romance fraud as documented by the EPC.

---

Validation Results

SynSEPA passed 41 of 42 automated validation checks covering:

• Basic dataset composition and size
• Fraud typology distribution
• Persona behavioural profiles
• Temporal patterns (hours, weekdays, time between transactions)
• Fraud typology signature verification
• Cross-border transaction rates
• Sequence integrity (ordering per account)
• Amount distribution shape

---

Intended Uses

Appropriate uses:

• Training and evaluating unsupervised or transformer based anomaly detection models
• Benchmarking generative models (VAE, GAN, Transformer, Diffusion) on tabular fraud data
• Sequence modelling research for financial fraud
• Academic research in payment fraud detection
• Developing and testing fraud detection pipelines without real customer data

Inappropriate uses:

• Training models intended for production deployment without further validation on real data
• Any use involving real customer data or real SEPA infrastructure
• Commercial fraud detection products without disclosure of synthetic data origin

---

Limitations

1. Synthetic data gap — All data is synthetic. Real fraud patterns may have nuances not captured in EPC typology descriptions.

2. No device/session signals — Real APP fraud detection also uses device fingerprints, session behaviour, and browser signals. SynSEPA covers only transaction-level signals.

3. Simplified IBAN structure — IBANs are syntactically correct but do not pass Mod-97 checksum validation.

4. Single year — Covers January-December 2024 only. Seasonal effects beyond this period are not represented.

5. Retiree cross-border rate — Marginally exceeds 10% target (achieved 10.3%) — a minor calibration artefact with no material impact on model training.

---

Citation

If you use SynSEPA in your research, please cite:

@dataset{synsep2025,
  title     = {SynSEPA: A Synthetic SEPA Instant Payment Dataset
               for APP Fraud Detection Research},
  author    = {Bajaj, Gaurav},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/EpiphanyTech/SynSEPA},
  note      = {Generated as part of the SEPAGen project.
               Calibrated against EBA-ECB 2025 Payment Fraud Report.}
}

---

References

1. EBA-ECB (2025). Joint Report on Payment Fraud. European Banking Authority / European Central Bank. December 2025.

2. EPC (2025). Payment Threats and Fraud Trends Report (EPC162-24 v2.0). European Payments Council. November 2025.

3. Lopez-Rojas, E. (2017). Synthetic Financial Datasets for Fraud Detection (PaySim). Kaggle.

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the dataset for any purpose, provided appropriate credit is given.

---

SynSEPA was generated entirely from synthetic data. No real customer data, real SEPA transactions, or real bank records were used in its creation.