andremessina/pix-fraud-br

PIX Fraud BR

A synthetic dataset of 2 million PIX instant payment transactions in the Brazilian financial system, designed for fraud detection research. Derived from the PaySim mobile money simulator and adapted to reflect Brazil's official PIX modalities, BCB regulatory rules, and real-world fraud patterns documented by Febraban.

Rows	2,000,000
Fraud cases	15,376 (0.77%)
Features	17
Real rows (PaySim TRANSFER, all 532,909 kept)	532,909
Synthetic rows	1,467,091
PaySim source	6,362,620 total rows — only TRANSFER (532,909) used
Period	January 2024 (BRT, UTC-3)

---

Background

PIX is Brazil's instant payment system, operated by the Banco Central do Brasil (BCB). Launched in November 2020, it processed over 63.8 billion transactions in 2024 — surpassing the combined volume of credit cards, debit cards, boleto, and TED. The rapid adoption has made it the primary target for social engineering fraud in Brazil, where fraudsters manipulate victims into authorizing transfers to mule accounts (contas laranja).

No public dataset of real PIX fraud transactions exists — financial institutions treat this data as proprietary. This dataset fills that gap by adapting PaySim's transaction simulator to match the PIX ecosystem.

---

Design Decisions

Transaction types

PaySim simulates five transaction types. Only TRANSFER was kept:

PaySim type	Decision	Reason
TRANSFER	Kept	P2P transfers between individuals — the only type with fraud in PaySim, and the primary PIX fraud vector
PAYMENT	Dropped	Recipients are merchant accounts with untracked balances (0/0 in 100% of cases) — pure noise
CASH-IN, CASH-OUT, DEBIT	Dropped	No direct PIX equivalent

PIX modalities

tipo_transacao follows BCB's Manual de Padrões para Iniciação do Pix v2.9.0, assigned with realistic P2P distributions:

Value	Description	Share
chave_pix	Transfer by PIX key (CPF, phone, email, random)	68%
dados_bancarios	Transfer by manual bank data (branch + account)	20%
pix_copia_e_cola	Copy-and-paste of QR Code payload	12%

Balance generation

PaySim's original balance columns track cumulative account state per time step, not the isolated before/after state of a single transaction — making them inconsistent in 98% of rows. All balance values were regenerated synthetically from valor_brl using distributions conditioned on fraud label:

Sender (pagador):

• Legitimate: k ~ LogNormal(1.5, 0.6) → median k ≈ 4.5x → pays ~22% of balance
• Fraud: k ~ LogNormal(0.05, 0.3) → median k ≈ 1.05x → empties ~95% of balance

Receiver (recebedor):

• Legitimate: LogNormal(10.0, 1.0) → median ≈ R$22k (normal account balance)
• Fraud (mule account): LogNormal(6.0, 1.5) → median ≈ R$400 (low pre-existing balance)

Consistency is enforced:

saldo_posterior_pagador   = saldo_anterior_pagador   − valor_brl
saldo_posterior_recebedor = saldo_anterior_recebedor + valor_brl

---

Schema

Column	Type	Description
id_pagador	string	Masked CPF of sender (***.XXX.XXX-**)
id_recebedor	string	Masked CPF of receiver (***.XXX.XXX-**)
tipo_transacao	string	PIX initiation modality (see above)
valor_brl	float	Transaction amount in BRL
saldo_anterior_pagador	float	Sender balance before transaction
saldo_posterior_pagador	float	Sender balance after transaction
saldo_anterior_recebedor	float	Receiver balance before transaction
saldo_posterior_recebedor	float	Receiver balance after transaction
datetime_brasilia	string	Transaction datetime in BRT (UTC-3)
hora_dia	int	Hour of day (0–23)
dia_semana	string	Day of week in Portuguese
dia_util	bool	True if Monday–Friday
horario_noturno	bool	True if 20h–06h (BCB default night period)
acima_limite_noturno	bool	True if valor_brl > R$1,000 and horario_noturno
razao_saldo_residual	float	saldo_posterior_pagador / saldo_anterior_pagador — proportion of balance remaining
proporcao_valor_recebedor	float	valor_brl / (valor_brl + saldo_anterior_recebedor) — transaction weight in receiver's total
fraude	int	Target — 1 = fraud, 0 = legitimate

---

Fraud Patterns

Fraud reflects the most common PIX attack in Brazil: social engineering or account takeover, followed by transfer to a conta laranja (mule account).

Signal	Legitimate	Fraud
razao_saldo_residual median	0.78 (paid 22%)	0.08 (emptied 92%)
saldo_anterior_recebedor median	R$22,152	R$440
saldo_anterior_recebedor IQR	R$11k – R$43k	R$151 – R$1,107

XGBoost validation (200k sample, scale_pos_weight)

Metric	Score
ROC-AUC	0.9950
PR-AUC	0.8646

Feature importances (top):

Feature	Importance	Interpretation
saldo_anterior_recebedor	68.1%	Mule accounts have ~50× lower balance than legitimate receivers
razao_saldo_residual	10.5%	Fraudulent senders nearly empty their accounts
saldo_anterior_pagador	5.7%	Supporting signal for sender profile
proporcao_valor_recebedor	3.1%	Transaction dominates mule account's total funds

---

Class Imbalance

The 0.77% fraud rate is intentional — realistic fraud rates in payment systems are low by design:

Dataset	Fraud rate
Real payment systems (BCB estimates)	0.01–0.1%
Kaggle Credit Card Fraud (benchmark)	0.17%
PIX Fraud BR	0.77%

Users should apply imbalanced learning techniques: class_weight='balanced', SMOTE, threshold tuning, and evaluate with PR-AUC rather than accuracy.

---

Synthetic Data Methodology

All 532,909 TRANSFER rows from PaySim (out of 6,362,620 total) were kept as the real base. 1,467,091 synthetic rows were generated via stratified bootstrap with Gaussian noise:

1. Rows are sampled with replacement, stratified by fraude
2. Gaussian noise (5% of each column's std) is added to valor_brl, saldo_anterior_pagador, saldo_anterior_recebedor
3. Posterior balances are re-derived to enforce consistency
4. Temporal features, risk ratios, and IDs are re-derived from noised base values
5. datetime_brasilia is shifted ±3 days to avoid exact duplicates

---

Known Limitations

• Transaction amounts: valor_brl inherits PaySim's distribution (median R$12k, mean R$191k), which is larger than typical Brazilian retail PIX transactions. This limits the discriminative power of proporcao_valor_recebedor.
• Temporal scope: Simulates a single month (January 2024). Does not capture seasonal patterns.
• Account IDs: Fully anonymized — cannot be used for graph-based (network) fraud detection.
• No multi-hop transfers: Real conta laranja networks often involve chains of 2–4 accounts. This dataset models only the first transfer.

---

Source & License

Derived from PaySim Synthetic Financial Dataset by Edgar Lopez-Rojas et al., licensed under CC BY-SA 4.0.

Regulatory references:

• BCB — Manual de Padrões para Iniciação do Pix v2.9.0
• BCB — Resolução BCB nº 6/2020 — Limites noturno
• Febraban — Golpes e fraudes com PIX (2024)
• Febraban — Autorregulação contas laranja (2024)

---

Citation

@dataset{messina2025pixfraudbr,
  author    = {Messina, André},
  title     = {PIX Fraud BR: Synthetic Brazilian PIX Fraud Detection Dataset},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/andremessina/pix-fraud-br},
  note      = {Derived from PaySim (Lopez-Rojas et al., 2016). Balance columns regenerated synthetically.}
}