Paper: arXiv 2512.12783

Authors: Atalay Denknalbant, Emre Sezdi, Zeki Furkan Kutlu, Polat Goktas

Abstract

Financial exclusion constrains entrepreneurship, increases income volatility, and widens wealth gaps. Underbanked consumers in Istanbul often have no bureau file because their earnings and payments flow through informal channels. To study how such borrowers can be evaluated we create a synthetic dataset of one hundred thousand Istanbul residents that reproduces first quarter 2025 TÜİK census marginals and telecom usage patterns. Retrieval augmented generation feeds these public statistics into the OpenAI o3 model, which synthesises realistic yet private records. Each profile contains seven socio demographic variables and nine alternative attributes that describe phone specifications, online shopping rhythm, subscription spend, car ownership, monthly rent, and a credit card flag. To test the impact of the alternative financial data CatBoost, LightGBM, and XGBoost are each trained in two versions. Demo models use only the socio demographic variables; Full models include both socio demographic and alternative attributes. Across five fold stratified validation the alternative block raises area under the curve by about one point three percentage and lifts balanced (F_{“1”}) from roughly 0.84 to 0.95, a fourteen percent gain. We contribute an open Istanbul 2025 Q1 synthetic dataset, a fully reproducible modeling pipeline, and empirical evidence that a concise set of behavioural attributes can approach bureau level discrimination power while serving borrowers who lack formal credit records. These findings give lenders and regulators a transparent blueprint for extending fair and safe credit access to the underbanked.

Complexity vs Empirical Score

  • Math Complexity: 2.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper uses standard machine learning models (CatBoost, LightGBM, XGBoost) with relatively simple implementation and presents clear empirical results (AUC, F1 scores) from a reproducible pipeline on a synthetic dataset, but does not involve advanced mathematical derivations or complex statistical theory.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Estimate credit risk for underbanked<br>using alternative data"] --> B["Data Synthesis:<br>Create Istanbul 2025 Q1 synthetic dataset<br>via RAG + OpenAI o3"]
  B --> C["Features & Splits:<br>Socio-demographic + Alternative attributes<br>5-fold stratified validation"]
  C --> D["Modeling:<br>Train CatBoost, LightGBM, XGBoost<br>(Demo vs Full models)"]
  D --> E["Outcomes:<br>AUC +1.3%, F1 0.84→0.95<br>Open dataset & pipeline released"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes an open synthetic Istanbul 2025 Q1 dataset and a fully reproducible modeling pipeline that uses alternative behavioural attributes for credit scoring.

Source sentence
We contribute an open Istanbul 2025 Q1 synthetic dataset, a fully reproducible modeling pipeline, and empirical evidence that a concise set of behavioural attributes can approach bureau level discrimination power while serving borrowers who lack formal credit records.

Assumes 3

  • The synthetic dataset accurately reproduces the 2025 Q1 TÜİK census marginals and telecom usage patterns of Istanbul residents. data
    Source sentence
    synthetic dataset of one hundred thousand Istanbul residents that reproduces first quarter 2025 TÜİK (TURKSTAT) census marginals and telecom usage patterns.
  • Alternative behavioural attributes are available and can be collected quickly with informed consent. execution
    Source sentence
    From an operational standpoint lenders must ask whether the data needed for the Full model can be gathered quickly and with informed consent.
  • Five‑fold stratified validation provides a reliable estimate of model performance. statistical
    Source sentence
    Across five fold stratified validation the alternative block raises area under the ROC curve by about 1.3 percentage points and lifts balanced F1 from roughly 0.84 to 0.95.

Data

  • Istanbul 2025 Q1 synthetic dataset synthetic public
    2025 Q1 · one hundred thousand Istanbul residents
    Source sentence
    synthetic dataset of one hundred thousand Istanbul residents that reproduces first quarter 2025 TÜİK (TURKSTAT) census marginals and telecom usage patterns.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Adding ten alternative attributes raises AUC by about 0.013 (1.3 percentage points) and improves balanced F1 from roughly 0.84 to 0.95, a 14% gain. predictive-accuracy
    Source sentence
    Across five fold stratified validation, the alternative block raises area under the ROC curve by about 1.3 percentage points and lifts balanced F1 from roughly 0.84 to 0.95, a 14 percent gain.
  • CatBoost Full flags 96 of 100 defaulters while issuing fewer than five false approvals, outperforming demographic‑only models. performance
    Source sentence
    CatBoost Full, for example, flags ninety six of one hundred defaulters while issuing fewer than five false approvals in the same batch, a separation level that pure demographic scorecards cannot reach.
  • A lender using LightGBM Full could approve roughly fourteen additional loans per hundred underbanked applicants without raising expected loss. performance
    Source sentence
    A lender that currently declines every applicant without a bureau file could adopt LightGBM Full, set a threshold that maintains the loss rate of its prime book, and approve roughly fourteen additional loans per hundred underbanked applicants.
  • Sensitivity analysis shows that lowering the cutoff by four AUC points still keeps expected loss below three percent, acceptable for most installment products. statistical
    Source sentence
    Sensitivity analysis in the notebook shows that lowering the cut off by four AUC points still keeps expected loss below three percent, a level acceptable for most installment products.
  • In 42% of cases where the prediction flips from reject to accept, changing a single variable (online shopping frequency, subscription cost, or phone age) is sufficient. other
    Source sentence
    In forty two percent of cases where the prediction flips from reject to accept, changing just one variable online shopping frequency, subscription cost, or phone age is sufficient.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 34 related papers). Coverage is partial, so this is neither replication nor refutation.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (29)

  1. The Global Findex Database 2021: Financial Inclusion, Digital Payments, and Resilience in the Age of COVID-19 (2022) — link
  2. Financial Inclusion and Inclusive Growth: A Review of Recent Evidence (2017) — link
  3. Invisible Primes: Fintech Lending with Alternative Data (2022) — link
  4. Behavior Revealed in Mobile Phone Usage Predicts Credit Repayment (2020) — link
  5. The Value of Big Data for Credit Scoring: Enhancing Financial Inclusion using Mobile Phone Data and Social Network Analytics (2019) — link
  6. Spatiotemporal Integration of Mobile, Satellite, and Public Geospatial Data for Enhanced Credit Scoring (2021) — link
  7. On the Rise of FinTechs: Credit Scoring Using Digital Footprints (2020) — link
  8. Super-app behavioral patterns in credit risk models: Financial, statistical and regulatory implications (2021) — link
and 21 more
  1. Deciphering Big Data in Consumer Credit Evaluation (2021) — link
  2. Does Psychometric Testing in Microfinance Actually Work?—The Case of Sogesol (2020) — link
  3. Psychometrics as a Tool to Improve Screening and Access to Credit (2015) — link
  4. Innovative Credit Risk Assessment: Leveraging Social Media Data for Inclusive Credit Scoring in Indonesia’s Fintech Sector (2025) — link
  5. Credit Scoring Using Machine Learning by Combing Social Network Information: Evidence from Peer-to-Peer Lending (2019) — link
  6. Judging Borrowers by the Company They Keep: Friendship Networks and Information Asymmetry in Online Peer-to-Peer Lending (2013) — link
  7. Credit Scoring with Social Network Data (2016) — link
  8. The Roles of Alternative Data and Machine Learning in Fintech Lending: Evidence from the LendingClub Consumer Platform (2019) — link
  9. Financial inclusion and fintech: a state-of-the-art systematic literature review (2025) — link
  10. The Critical Role of Alternative Datasets in Credit Assessment Using Machine Learning Techniques (2024) — link
  11. Predictive Analytics for Financial Inclusion: Using Machine Learning to Improve Credit Access for Under Banked Populations (2024) — link
  12. Enhancing Credit Scoring Accuracy with a Comprehensive Evaluation of Alternative Data (2024) — link
  13. Credit Score Classification Using Advanced Machine Learning: A Comprehensive Approach (2025) — link
  14. CatBoost: Gradient Boosting with Categorical Features Support (2017) — link
  15. LightGBM: A Highly Efficient Gradient Boosting Decision Tree (2017) — link
  16. A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection (1995)
  17. Unlocking Credit Access: Using Non-CDR Mobile Data to Enhance Credit Scoring for Financial Inclusion (2025) — link
  18. A Boosted Decision Tree Approach Using Bayesian Hyper-Parameter Optimization for Credit Scoring (2017) — link
  19. Comparing the Areas under Two or More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach (1988) — link
  20. XGBoost: A Scalable Tree Boosting System (2016) — link
  21. Explaining Machine Learning Classifiers through Diverse Counterfactual Explanations (2020) — link

0 of 29 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.