This dataset is fully synthetic — it contains no real Kenyan individuals, borrowers, accounts, or transactions. By requesting access you agree to (1) always disclose that the data is synthetic in any downstream use or publication, (2) not represent it as real Kenyan consumer or DCP portfolio data, and (3) not attempt to re-identify, link, or cross-reference any record with real individuals or institutions.

Kenya DCP Loan Performance & Collections Dataset (Synthetic) — v1.0

⚠️ This dataset is 100% synthetic. It does not contain, derive from, or represent any real Kenyan individual, business, loan, or mobile-money transaction. All identifiers (loan_id, borrower_id) are sequentially generated and carry no real-world meaning. Any resemblance between generated values and real records is coincidental.

1. Problem statement

Kenya has 600+ entities registered or licensed as Digital Credit Providers (DCPs) under the Central Bank of Kenya's 2022 regulatory regime. Unsecured, short-term digital lending is central to Kenyan financial inclusion but is also associated with loan stacking, opaque pricing, and high default/write-off rates. There is no public dataset that lets researchers, regulators, smaller DCPs, or academics build and benchmark loan default, collections, or responsible-lending models — real DCP portfolios are commercially confidential and CBK-held.

2. Why it matters in Kenya

Digital credit is often the first formal credit product for informal-sector and unbanked/underbanked Kenyans.
Loan stacking (borrowing from multiple apps simultaneously) is a documented systemic risk that conventional bureau data captures poorly.
CBK's DCP regulations require risk monitoring and responsible-lending practices, but smaller/newer providers and researchers have no data to prototype against.
Collections practices in Kenyan digital lending have drawn regulatory and consumer- protection scrutiny — a labeled collections dataset supports research into fairer, more effective collections strategies.

3. Existing data gap

Public Kenyan fintech datasets are almost all either (a) generic global credit-scoring toy datasets with no Kenyan context, or (b) small survey-based financial-inclusion datasets (e.g., FinAccess) that lack loan-level granularity and outcome labels. No open dataset combines loan-level terms, borrower behavioral proxies, loan-stacking indicators, collections actions, and repayment outcomes in a Kenyan DCP context.

4. Dataset description

Grain: one row per loan (a borrower may appear across multiple loans — 3,521 synthetic borrowers → 5,000 loans, avg. 1.42 loans/borrower, matching realistic repeat-borrowing patterns).
Size: 5,000 records. Splits: train 3,499 / val 749 / test 753 (≈70/15/15, stratified by repayment_status).
Time span (synthetic): applications dated Jan 2024–Dec 2025.
Files:
- kenya_dcp_loan_dataset_full.csv — all 5,000 records with split column
- kenya_dcp_loan_dataset_train.csv / _val.csv / _test.csv — pre-split, no split column

5. Schema

Column	Type	Description
`loan_id`	string	Synthetic unique loan identifier
`borrower_id`	string	Synthetic unique borrower identifier (repeats across loans)
`prior_loans_count`	int	Number of previous loans by this borrower in the dataset
`application_date`, `disbursement_date`, `due_date`	date	Loan lifecycle dates
`lender_type`	categorical	DCP category (mobile-only app, bank-backed, SACCO digital, fintech-MFI hybrid)
`loan_channel`	categorical	Mobile App / USSD / Agent-assisted / Bank digital channel
`loan_purpose`	categorical (5% missing)	Self-reported purpose
`loan_amount_kes`	float	Principal disbursed, KES
`loan_term_days`	int	7/14/30/60/90-day term
`fee_pct`	float	Flat facilitation fee, % of principal (Kenyan digital-lender norm, not amortized interest)
`principal_plus_fee_kes`	float	Total amount owed at term end
`age`, `gender`, `county`, `urban_rural`, `employment_type` (4% missing)	—	Borrower demographics
`estimated_monthly_income_kes`	float (9% missing)	Self-reported/estimated income
`device_type`	categorical	Smartphone / feature phone (USSD-only)
`mobile_money_tenure_months`	int	Proxy for digital financial footprint
`avg_monthly_mm_inflow_kes`, `avg_monthly_mm_txn_count`	float/int	Mobile-money activity proxies (not real M-Pesa data)
`num_active_loan_apps_last_90d`	int	Loan-stacking proxy
`prior_default_count`	int	Defaults on earlier loans in this synthetic history
`credit_score_alt`	float (300–850)	Synthetic alternative credit score
`restructured_flag`	binary	Loan previously renegotiated
`repayment_channel`	categorical	M-Pesa STK Push / Agent cash / Bank transfer / Airtel Money
`number_of_reminders_sent`, `number_of_collection_calls`	int	Collections effort
`days_to_first_repayment_attempt`	int	Behavioral timing signal
`days_past_due`	int	DPD at 120-day observation window
`repayment_status`	categorical	Paid on time / Paid late / Defaulted — primary label
`default_flag`	binary	1 = Defaulted — primary binary target
`written_off_flag`	binary	Default with DPD > 90
`amount_recovered_kes`, `recovery_rate`	float	Collections/recovery outcome
`is_synthetic`	bool	Always `True`
`dataset_version`	string	`kenya_dcp_synth_v1.0`

6. Generation methodology

Programmatic borrower simulation (numpy/pandas): demographics, county (weighted toward real population concentration, e.g. Nairobi/Kiambu/Mombasa), employment type, income (log-normal, conditioned on employment + urban/rural), mobile-money tenure and inflow proxies.
Loan-term simulation: loan amount as a function of income, prior-loan graduation, and lender type (bank-backed lenders issue larger loans than mobile-only apps); fee structure reflects real Kenyan flat-fee digital-lending pricing conventions (higher % for shorter terms).
Outcome simulation via logistic risk model (not independent random draws): default probability is a function of loan-to-income ratio, loan-stacking, prior defaults, mobile-money tenure, employment informality, restructuring history, and alt credit score, plus idiosyncratic noise (so the signal is realistic, not perfectly separable). A second logistic model determines "paid late" vs "on time" among non-defaults, reduced by collections effort (reminders/calls) — modeling diminishing-returns collections behavior.
Recovery simulation: partial, DPD-dependent, collections-effort-boosted recovery for defaults; full recovery for on-time/late loans.
Missingness injection: realistic missingness on self-reported fields (income 9%, employment type 4%, loan purpose 5%, age 2%) — mimics real DCP data-quality issues.
No LLM-generated free text in this release (structured/tabular only); reasoning above (risk-factor weighting, vocabulary design, realistic fee/term conventions) was informed by domain knowledge of Kenyan digital lending rather than external real data.
Fully reproducible: fixed seed (20260908), deterministic pipeline, script-based — parameters documented above.

7. Class distribution

Split	Paid on time	Paid late	Defaulted
Train	65.4%	23.1%	11.5%
Val	65.5%	23.0%	11.5%
Test	65.3%	23.1%	11.6%

Overall default rate 11.5% — calibrated to published ranges for unsecured short-term Kenyan digital credit, and intentionally imbalanced-but-workable (not degenerate).

8. Quality control performed

13 automated checks, all passing:

No duplicate loan_ids or fully duplicate rows
No invalid values (negative amounts, out-of-range ages, recovery rate outside [0,1])
Due date always after disbursement date
Default rate within a realistic band (5–25%)
loan_id sequence uncorrelated with target (rules out ordinal leakage, |r| = 0.019)
No phone-number-like or National-ID-like strings anywhere in the data
Missingness rates within designed bands
Loan amount correlates with income (r = 0.72) but is not deterministic (i.e., not a disguised formula the model could trivially invert)
County / urban-rural geographic consistency

Baseline model sanity benchmark (see below) confirms the dataset is learnable (meaningfully above random) without being trivially solvable — the profile you want in a benchmark dataset.

Model	AUC-ROC	PR-AUC
Logistic Regression (balanced)	0.676	0.247
LightGBM	0.622	0.166

(Simple linear model outperforming a default-tuned GBM here is itself a realistic and useful benchmarking signal — it means the risk relationships are mostly monotonic/linear by design, and a well-tuned GBM should be able to match or exceed logistic regression with proper regularization/tuning — a good exercise for users of the benchmark.)

Top predictive features (LightGBM importance): fee_pct, credit_score_alt, county, avg_monthly_mm_inflow_kes, mobile_money_tenure_months, estimated_monthly_income_kes, age, loan_amount_kes — all plausible, pre-outcome, non-leaking features.

9. Bias considerations

County weights approximate population concentration but are illustrative, not census- accurate — do not use this dataset to make real geographic policy claims.
Income-by-employment-type base rates are directional approximations, not official wage statistics.
Users building fairness-aware models should treat gender, county, and urban_rural as protected/sensitive attributes and test for disparate impact before any real-world-adjacent use (e.g., as a teaching example for fairness audits).

10. Privacy and ethical considerations

No real people, accounts, phone numbers, national IDs, or M-Pesa transaction codes.
All identifiers are synthetic sequence strings (DCP-KE-######, BOR-######) with formats deliberately distinct from real Safaricom/CBK ID conventions.
Dataset must not be represented as real DCP portfolio data in any downstream publication — always disclose as synthetic (per is_synthetic=True field).

11. Licensing

Released for research, educational, and product-prototyping use. Treat as CC0-equivalent synthetic data — attribution appreciated but not required. Not for use in claiming regulatory compliance testing without disclosure that underlying data is synthetic.

12. Recommended uses

Loan default / credit-risk model prototyping (default_flag, repayment_status)
Collections-strategy optimization (number_of_collection_calls vs. recovery outcomes)
Alternative credit scoring research (mobile-money-proxy features vs. credit_score_alt)
Loan-stacking / responsible-lending research (num_active_loan_apps_last_90d)
Teaching material for fairness/bias audits in credit models
Baseline for a future "Kenyan Digital Credit Benchmark" if paired with additional real or synthetic datasets (fraud, SME cash-flow, etc.)

13. Recommended evaluation benchmarks

Default prediction: AUC-ROC, PR-AUC (primary, given class imbalance), Brier score (calibration), KS-statistic (industry-standard for credit scoring)
Multi-class repayment status: macro-F1, confusion matrix
Recovery-rate regression: MAE, R²
Recommended baseline models: Logistic Regression, LightGBM/XGBoost, and a simple MLP; report performance stratified by lender_type and urban_rural to surface fairness gaps.

Downloads last month: -

Total file size:

3.03 MB