README.md

metadata

annotations_creators:
  - machine-generated
language:
  - en
license:
  - mit
multilinguality:
  - monolingual
pretty_name: Electric Sheep Alternative Credit Data for Thin-File Users
size_categories:
  - 100<n<1000
source_datasets:
  - original
tags:
  - credit-scoring
  - financial-inclusion
  - thin-file
  - synthetic-data
  - behavioral-finance
  - alternative-data
  - tabular
  - fintech
  - machine-learning
  - synthetic
task_categories:
  - tabular-classification
  - tabular-regression
configs:
  - config_name: US
    data_files: US/*.parquet
  - config_name: NG
    data_files: NG/*.parquet
  - config_name: IN
    data_files: IN/*.parquet
  - config_name: BR
    data_files: BR/*.parquet
data_type: synthetic

⚠️ Synthetic dataset — Parameterized from published SSA literature, not real observations. Not suitable for empirical analysis or policy inference.

Electric Sheep — Alternative Credit Data for Thin-File Users

Dataset Summary

The Electric Sheep Alternative Credit Dataset is a synthetic dataset for modeling creditworthiness using behavioral financial data rather than traditional credit history. It targets thin-file users — individuals excluded from conventional credit scoring due to insufficient loan or credit card history.

Four culturally authentic regions:

Region	Currency	Dominant Channel	Median Income
🇺🇸 US (United States)	USD	POS / Bank	~$17,400
🇳🇬 NG (Nigeria)	NGN	Mobile Money	~₦2,400
🇮🇳 IN (India)	INR	UPI	~₹1,900
🇧🇷 BR (Brazil)	BRL	PIX / Bank	~R$3,500

Files Per Region

Each region contains 2 files in {region}/:

File	Description
credit_profiles.parquet	User-level profiles with demographics, 19 behavioral features, and credit labels (1 row per user, 30 columns)
metadata.json	Generation statistics

Raw transactions are in transactions/{region}.parquet (separate download, ~900 transactions per user).

Loading the Dataset

from datasets import load_dataset

# Load US region
ds = load_dataset("electricsheepafrica/electric-sheep-credit", name="US")
profiles = ds["train"]  # credit_profiles.parquet
print(profiles[0])

Or with pandas:

import pandas as pd

profiles = pd.read_parquet("US/credit_profiles.parquet")
transactions = pd.read_parquet("US/transactions.parquet")

Credit Profiles Schema (30 columns)

Demographics:

• user_id — Unique identifier (US_000000)
• age_range — 18-25, 26-35, 36-45, 46-55, 55+
• income_type — salary, gig, business, unemployed
• region — US, NG, IN, BR
• currency — USD, NGN, INR, BRL
• account_tenure_days — Length of financial activity
• primary_archetype — Behavioral pattern (stable_salaried, gig_worker, etc.)
• secondary_archetype — Secondary pattern

Behavioral Features (19):

• avg_monthly_income — Mean monthly inflow
• income_volatility — Coefficient of variation of income
• income_frequency — Income events per month
• income_trend — Linear trend (positive = growing)
• income_gap_months — Months with zero income
• avg_monthly_spend — Mean monthly outflow
• spending_volatility — Variability of spending
• essential_spend_ratio — Fraction on food, transport, bills, healthcare
• discretionary_spend_ratio — Fraction on non-essentials
• betting_ratio — Fraction on betting/gambling
• avg_balance — Mean balance
• min_balance — Lowest balance reached
• max_balance — Highest balance reached
• overdraft_frequency — Fraction of transactions with negative balance
• bill_payment_consistency — Regularity of bills (0–1)
• recurring_expense_ratio — Fraction of spending that recurs
• transaction_frequency — Transactions per day
• cashflow_stability_score — Composite stability score (0–1)
• risk_behavior_score — Composite risk indicator (0–1)

Credit Labels:

• credit_outcome — good, bad, indeterminate
• default_probability — Estimated default risk (0–1)
• risk_bucket — low, medium, high

Transactions Schema

Column	Type	Description
transaction_id	string	UUID
user_id	string	Foreign key
timestamp	datetime	Transaction time
amount	float	Signed: +income, -expense
transaction_type	string	credit or debit
category	string	food, transport, bills, entertainment, betting, transfer, savings, healthcare, other
channel	string	bank, cash, POS, mobile_money
merchant_type	string	Region-specific merchant
balance_estimate	float	Running balance
is_recurring	bool	Recurring payment flag
counterparty	string	Transfer partner (if applicable)
corridor	string	Remittance path (e.g., US→NG)

Feature Correlations with Creditworthiness

Feature	US	NG	IN	BR	Interpretation
overdraft_frequency	−0.73***	−0.84***	−0.80***	−0.74***	Strongest negative
cashflow_stability_score	+0.58***	+0.70***	+0.71***	+0.66***	Strongest positive
betting_ratio	−0.36*	−0.60***	−0.65***	−0.60***	Significant negative
avg_balance	+0.49**	+0.56**	+0.40*	+0.48**	Liquidity buffer
bill_payment_consistency	+0.50**	+0.50**	+0.38*	+0.37*	Reliability signal

Label Distribution (V2 — Simulated Loans)

Region	Good	Bad	Indeterminate	Default Rate
US	63%	20%	17%	0.22
NG	47%	37%	16%	0.32
IN	60%	30%	10%	0.25
BR	47%	33%	20%	0.30

Use Cases

• Credit scoring model development for thin-file populations
• Benchmarking alternative data approaches
• Financial inclusion research
• Fairness and bias testing
• Cross-regional behavioral finance studies

Limitations

• Synthetic data (not from real financial institutions)
• Simulated credit outcomes
• Simplified temporal patterns
• Approximate regional income levels

Citation

@dataset{electric_sheep_2026,
  title={Electric Sheep: Alternative Credit Data for Thin-File Users},
  author={ElectricSheepAfrica},
  year={2026},
  license={MIT},
  url={https://huggingface.co/datasets/electricsheepafrica/electric-sheep-credit}
}

License

MIT — free for research and commercial use.