azizstark/synthetic-chargeback-cases

Synthetic Chargeback Cases for Representment Win Prediction

Dataset Description

A synthetic dataset of 10,000 credit card chargeback cases designed for training and evaluating machine learning models that predict representment win probability — the likelihood a merchant will win if they fight a chargeback dispute.

The dataset models realistic chargeback workflows including evidence collection, reason code categorization, customer/merchant profiling, and outcome determination — suitable for building AI-assisted analyst tools that summarize cases, assemble evidence, and provide policy-aligned recommendations (REPRESENT / REFUND / ESCALATE).

Supported Tasks

• Binary Classification: Predict whether a merchant will win (representment_won = 1) or lose (representment_won = 0) a chargeback representment.
• Probability Estimation: Estimate calibrated win probability for prioritizing cases.
• Explainability Research: Analyze which evidence types and case features drive outcomes using SHAP or similar methods.

Languages

English (all text fields are in English).

Dataset Summary

Property	Value
Total cases	10,000
Represented cases (with outcome)	7,804
Accepted cases (no outcome)	2,196
Date range	Jan 2024 – Dec 2025
Card networks	Visa (65.6%), Mastercard (34.4%)
Reason codes	15 (Visa + Mastercard)
Chargeback categories	4 (Fraud, Consumer Dispute, Authorization, Processing Error)
Total columns	65
Overall win rate (represented)	~69.7%
Transaction amount range	$10.00 – $6,000.00
Currencies	USD (82.6%), EUR (9.3%), GBP (4.8%), CAD (3.3%)

Dataset Structure

Data Splits

The dataset is split by time (not randomly) to simulate real-world deployment where models predict future cases from past data:

Split	File	Rows	Time Period	Purpose
train	chargeback_cases_train.csv	4,682	Jan 2024 – ~Jul 2025	Model training
validation	chargeback_cases_val.csv	1,561	~Aug 2025 – ~Oct 2025	Hyperparameter tuning
test	chargeback_cases_test.csv	1,561	~Nov 2025 – Dec 2025	Final evaluation
full	chargeback_cases_full.csv	10,000	Jan 2024 – Dec 2025	All cases (including non-represented)

Note: The train/val/test splits contain only represented cases (where case_disposition == "REPRESENTED") and have a binary representment_won label. The full dataset also includes accepted cases (where the merchant did not fight the chargeback), which have NaN for representment_won.

Additional Files

File	Description
reason_codes_reference.csv	Lookup table for all 15 Visa/Mastercard reason codes with categories, descriptions, base win rates, and required evidence

Data Fields

The dataset contains 65 columns organized into the following groups:

Identifiers & Dates (5 columns)

Column	Type	Description
case_id	string	Unique case identifier (e.g., CB-2024-00001)
transaction_id	string	Unique transaction identifier (e.g., TXN-550639328)
transaction_date	date	Date of the original transaction
dispute_date	date	Date the chargeback was filed
resolution_date	date	Date the case was resolved

Transaction Features (10 columns)

Column	Type	Description
transaction_amount	float	Dollar amount of the disputed transaction ($10–$6,000)
amount_log	float	Log-transformed amount (log1p(amount))
amount_vs_merchant_avg	float	Ratio of transaction amount to merchant's average ticket
currency	string	Transaction currency (USD, EUR, GBP, CAD)
payment_method	string	Payment type (credit, debit, prepaid)
is_recurring	binary	1 if subscription/recurring billing
is_card_present	binary	1 if physical card was used (card-present transaction)
is_cross_border	binary	1 if transaction crossed country borders
is_digital_goods	binary	1 if product is digital (SaaS, streaming, downloads)
card_network	string	Visa or Mastercard

Reason Code Features (5 columns)

Column	Type	Description
reason_code	string	Network-specific reason code (e.g., Visa_10.4, MC_4837)
reason_category	string	High-level category: FRAUD, CONSUMER_DISPUTE, AUTHORIZATION, PROCESSING_ERROR
reason_description	string	Human-readable description of the reason code
reason_code_encoded	int	Integer-encoded reason code (0–14)
category_encoded	int	Integer-encoded category (0–3)

Customer Features (9 columns)

Column	Type	Description
customer_id	string	Anonymized customer identifier
customer_tenure_months	int	Months the customer has been with the bank
customer_total_orders	int	Total historical orders with the merchant
customer_prior_disputes	int	Number of previous chargebacks filed
customer_prior_win_rate	float	Win rate on past disputes (0–1)
customer_risk_score	float	Composite risk score (0–1, higher = riskier)
customer_account_verified	binary	1 if KYC-verified account
customer_email_domain	string	Email provider domain
is_repeat_disputer	binary	1 if customer has filed 2+ previous disputes

Merchant Features (8 columns)

Column	Type	Description
merchant_id	string	Anonymized merchant identifier
merchant_name	string	Synthetic merchant name
merchant_category_code	int	MCC code (e.g., 5732 = Electronics)
merchant_category_name	string	Human-readable MCC name
merchant_monthly_volume	int	Monthly transaction volume
merchant_chargeback_rate	float	Chargeback-to-transaction ratio (%)
merchant_avg_ticket	float	Average transaction amount for this merchant
merchant_years_active	int	Years the merchant has been processing payments

Evidence Features (16 columns)

Column	Type	Description
has_delivery_tracking	binary	1 if shipment tracking number on file
has_delivery_confirmation	binary	1 if delivery confirmed by carrier
has_delivery_signature	binary	1 if signed proof of delivery exists
has_avs_match	binary	1 if billing address matched (Address Verification System)
has_cvv_match	binary	1 if CVV security code matched at authorization
has_3ds_authentication	binary	1 if 3D Secure authentication was completed
has_customer_communication	binary	1 if emails/chat with customer are on file
has_refund_policy_shown	binary	1 if refund policy was displayed at checkout
has_terms_accepted	binary	1 if customer accepted Terms & Conditions
has_ip_geomatch	binary	1 if customer IP location matches billing address
has_device_fingerprint	binary	1 if device recognized from previous sessions
has_usage_proof	binary	1 if login/access logs show product was used
has_prior_non_disputed_txn	binary	1 if customer had previous successful orders
evidence_count	int	Total count of evidence items on file
evidence_completeness	float	Fraction of required evidence items present (0–1)
evidence_strength_score	float	Weighted composite evidence score (0–1)

Timing Features (5 columns)

Column	Type	Description
days_to_dispute	int	Days between transaction and chargeback filing
days_to_respond	int	Days the merchant took to respond
deadline_days	int	Total days allowed by network (Visa=30, MC=45)
response_within_deadline	binary	1 if merchant responded within the deadline
is_holiday_period	binary	1 if transaction occurred during holiday season (Nov–Jan 7)

Derived Features (2 columns)

Column	Type	Description
dispute_velocity	float	Rate of disputes per month of tenure
mcc_encoded	int	Integer-encoded merchant category code

Target & Outcome (3 columns)

Column	Type	Description
win_probability_true	float	Ground-truth win probability used to generate the binary outcome
case_disposition	string	REPRESENTED (merchant fought) or ACCEPTED (merchant accepted loss)
representment_won	binary	Target variable — 1 if merchant won the representment, 0 if lost, NaN if not represented

Reason Codes

The dataset covers 15 real Visa and Mastercard reason codes across 4 categories:

Category	Code	Description	Base Win Rate	Frequency
FRAUD	Visa_10.4	Fraud — Card Not Present	32%	25.4%
FRAUD	Visa_10.5	Fraud — Counterfeit Transaction	18%	3.1%
FRAUD	MC_4837	No Cardholder Authorization	28%	8.3%
FRAUD	MC_4863	Cardholder Does Not Recognize	38%	3.9%
CONSUMER_DISPUTE	Visa_13.1	Merchandise/Services Not Received	55%	14.6%
CONSUMER_DISPUTE	Visa_13.3	Not as Described / Defective	42%	7.9%
CONSUMER_DISPUTE	MC_4853	Goods/Services Not as Described	40%	7.4%
CONSUMER_DISPUTE	MC_4855	Goods/Services Not Received	56%	4.9%
AUTHORIZATION	Visa_11.1	Card Recovery Bulletin	22%	2.8%
AUTHORIZATION	Visa_11.2	Declined Authorization	52%	4.9%
AUTHORIZATION	MC_4808	Authorization Related	48%	7.1%
PROCESSING_ERROR	Visa_12.1	Late Presentment	28%	2.9%
PROCESSING_ERROR	Visa_12.2	Incorrect Transaction Code	62%	2.0%
PROCESSING_ERROR	Visa_12.5	Incorrect Amount	68%	2.0%
PROCESSING_ERROR	MC_4834	Duplicate Processing	72%	2.9%

Dataset Creation

Generation Process

The dataset was generated synthetically using a Python script (generate_dataset.py) with the following methodology:

1. Customer Profiles: 3,000 unique customers with realistic tenure, order history, dispute history, and risk profiles drawn from exponential and beta distributions.
2. Merchant Pool: 20 merchants across 14 MCC categories with varying volumes, chargeback rates, and ticket sizes.
3. Transaction Generation: Each case is assigned a reason code (weighted by real-world frequency), merchant, customer, and transaction details.
4. Evidence Simulation: Evidence flags are generated with base rates that vary by reason category, transaction amount, digital vs. physical goods, and customer tenure — reflecting real-world correlations.
5. Outcome Generation: Win probability is computed from a rule-based function incorporating reason code base rates, evidence presence, customer history, amount, and timing. Binary outcomes are sampled from this probability with added noise.
6. Time-Based Splitting: Represented cases are sorted by transaction date and split 60/20/20 into train/val/test.

Realistic Correlations Built In

• 3D Secure authentication provides a large boost for fraud cases (liability shift)
• Delivery confirmation + signature are critical for consumer dispute cases
• Higher transaction amounts reduce win probability (harder to defend)
• Cross-border transactions carry a penalty
• Repeat disputers (serial filers) are easier to defend against
• Digital goods require usage proof instead of delivery evidence
• Missing deadline response results in automatic loss

Random Seeds

All random number generators use seed=42 for full reproducibility.

Intended Use

Primary Use Cases

• ML model training: Train classifiers to predict chargeback representment outcomes
• Explainability research: Study feature importance and SHAP explanations in financial decision-making
• AI assistant development: Build AI-assisted analyst tools for chargeback review
• Educational: Learn about chargeback processes, evidence requirements, and ML applied to fintech

Out-of-Scope Uses

• Production fraud detection: This is synthetic data and should not be used as a substitute for real transaction data in production fraud systems.
• Regulatory compliance: The dataset does not reflect actual card network rules or regulatory requirements.

Baseline Model Performance

An XGBoost classifier trained on this dataset achieves:

Metric	Value
AUC-ROC	0.719
Accuracy	70.6%
Precision	73.2%
Recall	90.9%
F1 Score	0.811
Brier Score	0.188

The model uses 45 features (numeric, binary, and encoded categorical) with Platt scaling (sigmoid calibration) for probability calibration.

Limitations and Biases

• Synthetic data: All cases are generated programmatically. While correlations are designed to be realistic, they may not capture the full complexity of real-world chargeback patterns.
• Simplified evidence model: Real evidence evaluation involves document quality, completeness of narratives, and network-specific rules that are not captured here.
• Fixed merchant pool: Only 20 merchants are simulated, which limits merchant-level diversity.
• No temporal drift: The data generation process is stationary — real-world chargeback patterns shift over time due to policy changes, fraud trends, and seasonal effects.
• US-centric: Amount distributions and merchant categories are primarily modeled on US market patterns.

Citation

If you use this dataset in your research or projects, please cite:

@misc{synthetic_chargeback_cases_2025,
  title={Synthetic Chargeback Cases for Representment Win Prediction},
  author={azizstark},
  year={2025},
  url={https://huggingface.co/datasets/azizstark/synthetic-chargeback-cases},
  note={Synthetic dataset of 10,000 credit card chargeback cases for representment win prediction}
}

License

This dataset is released under the MIT License.

Dataset Card Contact

For questions or feedback, please open an issue on the dataset's discussion tab.