THESIS

Explainable AI for Financial Inclusion

An explainable, fairness-audited credit-scoring framework for the 1.4 billion adults locked out of formal credit.

MSC DATA SCIENCE & BUSINESS INTELLIGENCE · EDC PARIS BUSINESS SCHOOL

Read the full thesis (PDF)

1.4 billion adults worldwide lack access to formal credit, not because they can't repay, but because traditional credit scoring requires a bureau history that thin-file and no-file populations don't have. This thesis asks whether machine learning, made explainable via SHAP and audited for fairness, can begin to close that gap without breaking the transparency and robustness requirements the EU AI Act now imposes on high-risk AI systems like credit scoring.

I built and validated the Explainable Inclusive Credit Scoring (EICS) Framework: a three-pillar pipeline (gradient-boosted prediction, SHAP-based explanation, fairness-audited threshold calibration) tested on 1.35 million real Lending Club loan records and replicated on a second, structurally different dataset (Home Credit, 307,511 records) to confirm it wasn't a fluke of one lender's data.

The framework does not claim thin-file populations are direct beneficiaries yet. The datasets available for this thesis, by construction, only include borrowers who already have a bureau credit history, so that claim can't be tested directly here. What the thesis does show is that the framework is built, empirically validated twice, and ready for that deployment as the next step.

The problem

The cause is informational, not behavioural: a young entrepreneur with a viable business but no bureau credit history is invisible to a scoring model trained on bureau data, and so is a salaried worker paid in cash in an informal economy. Two developments made a different approach possible. Gradient-boosted ML models now outperform the logistic regression that has been the credit-scoring industry standard for decades, and explainable-AI techniques like SHAP can decompose a model's predictions into interpretable contributions, removing the black-box objection that has historically kept ML out of regulated lending. The EU AI Act adds the regulatory push: Article 86 gives affected persons a right to a clear, meaningful explanation of an automated credit decision.

What I built: the EICS Framework

Three pillars, plus two compliance layers. Prediction: a gradient-boosting classifier (XGBoost or LightGBM, empirically interchangeable) trained on a SMOTE-balanced sample with Bayesian hyperparameter search. Explanation: TreeSHAP computes both a global feature-importance ranking and per-application local explanations, which I translate into structured, plain-language decision narratives, not raw SHAP numbers. Fairness: an audit across three subgroup dimensions (income quartile, home ownership, verification status), with group-specific decision-threshold calibration to close the gaps the audit finds. Compliance Layer A operationalises the EU AI Act's Article 13 transparency requirement through those decision narratives; Compliance Layer B operationalises Article 15 robustness through fixed random seeds, versioned artefacts, and the cross-dataset replication described below.

What the models actually showed

Gradient boosting (LightGBM AUC-ROC 0.7293, XGBoost 0.7278) beat logistic regression (0.7158) by a statistically significant but modest margin. That margin is deliberately honest: I removed 50 features that leak information from after a loan originates, which is what lets other published analyses of this same dataset report inflated AUCs of 0.85 to 0.90. The 0.73 figure is the ceiling for predicting default at the moment someone actually applies. Random Forest, evaluated with reasonable default hyperparameters rather than the same tuning budget as the gradient boosters, came out essentially tied with logistic regression; whether a fully tuned Random Forest would close that gap is a question this thesis leaves open, not one it answers.

What drives the model's decisions

Lending Club's own assigned loan grade is the single strongest predictor by a wide margin. Behind it, seven of my engineered features made the top 20 SHAP ranking, including a numeric encoding of loan sub-grade at rank 2, but they complement the underlying bureau signal rather than replacing it. The case I find most interesting in the whole analysis is a high-income borrower who actually defaulted: a grade-and-income heuristic would have read them as low risk, but the SHAP breakdown surfaced the real signal underneath, a high installment-to-income ratio, heavy revolving-credit utilisation, and a recent delinquency flag. That is the actual argument for explainable ML in credit scoring: not that it is more accurate in the abstract, but that it can catch risk a simpler heuristic misses, and show its work while doing it.

Is it fair?

Before any correction, all three subgroup comparisons failed the four-fifths fairness rule. Low-income borrowers were predicted to default at roughly 3.2 times the rate of the highest-income quartile; renters at roughly 3.8 times the rate of owners. Group-specific threshold calibration, tuned per subgroup pair rather than applying one global threshold, cut the fairness gap (Equalised Odds Difference) by 97 to 99 percent across all three comparisons. The home-ownership calibration did this while also more than doubling the model's F1 score, fairness and accuracy moved in the same direction, because the uncalibrated model's default threshold of 0.5 was never a good operating point to begin with on this imbalanced data.

Does it hold up on a different dataset?

I replicated the full pipeline on Home Credit Default Risk, a dataset from a different lender serving different markets (307,511 applications, an 8 percent default rate versus Lending Club's 20 percent, and an entirely different feature taxonomy), using the exact hyperparameters tuned on Lending Club, with no re-tuning. AUC-ROC held: 0.7532, actually slightly higher than on Lending Club. Not one SHAP feature name overlapped between the two datasets, because they don't share column names, but the pattern was identical: both models leaned hardest on externally-sourced credit-bureau-style scores as their top predictors. The architecture transfers to a structurally different lending context without modification.

What this doesn't show yet

Every applicant in both datasets already has a formal credit-bureau footprint. That means this thesis validates that the EICS Framework works: it predicts, explains, and can be audited and calibrated for fairness. It does not yet validate the financial-inclusion claim that motivated the project, because the data available doesn't include the thin-file and no-file population that claim is actually about. The framework is the bridge; walking across it, testing it against a population it was built to eventually serve, is the next study, done in partnership with an institution that actually lends into that population.

FIGURES

The findings, in pictures

A curated set of figures from the thesis, captioned in plain language. The full figure set, methodology, and appendices are in the PDF above.

Bar chart and pie chart showing 1,076,751 non-defaults versus 268,559 defaults in the Lending Club dataset
The dataset: 1.35 million Lending Club loans, about 20% ended in default. I trained on a 500,000-loan stratified sample that preserves that same rate exactly.
Bar chart showing default rate rising monotonically from 6% at Grade A to 49.9% at Grade G
Lending Club's own risk grade works, cleanly: default rate rises from 6% at Grade A to 50% at Grade G in a straight line. That's why grade turns out to be the single strongest predictor later on.
ROC curves for Logistic Regression, Random Forest, XGBoost, and LightGBM
The four models head-to-head. LightGBM and XGBoost edge out logistic regression and Random Forest, but deliberately by a modest margin (0.729 vs 0.716 AUC): I stripped out 50 features that leak information from after a loan is issued, which is what lets other published analyses of this dataset report inflated 0.85+ AUCs. This is the honest number for predicting default at the moment someone applies.
Horizontal bar chart of the top 20 features by mean absolute SHAP value, led by grade
What the best model actually leans on, ranked. Lending Club's own loan grade dominates. Seven of my engineered features made the top 20, but they add to the bureau signal, they don't replace it.
SHAP beeswarm plot showing the direction and magnitude of each feature's effect across test instances
Same ranking, but showing direction, not just magnitude. Each dot is one loan. Red means a high feature value; its position shows whether that pushed the prediction toward default (right) or away from it (left).
SHAP dependence plot for loan grade, showing a monotonic relationship with SHAP value
Zooming into the single strongest feature: the relationship is clean and monotonic, each step down in grade increases the model's risk estimate in a straight line, with no reversals that would suggest the model is picking up noise.
SHAP waterfall plot for a clear high-risk applicant, predicted default probability 0.771
One real application, decomposed. This borrower's own risk grade alone pushed the predicted default probability up sharply; the model classified them high-risk at 77%, and they did default.
SHAP waterfall plot for a high-income borrower with high predicted default risk, who actually defaulted
The case I find most interesting in the whole thesis. This borrower had high income, which a grade-and-income check would read as low risk. SHAP surfaced the real signal instead: a high installment-to-income ratio, heavy revolving-credit use, a recent delinquency flag. They defaulted. Income alone would have missed this.
Bar charts comparing DPR, EOD, and DIR fairness metrics before and after threshold calibration, across three subgroup pairings
Before any correction (grey), the model was unfair by a wide margin across income, home ownership, and verification status. After I calibrated the decision threshold separately per subgroup (green), the fairness gap collapsed by 97 to 99% across the board.
Line chart showing Equalised Odds Difference and overall F1-score as a function of the classification threshold applied to the lowest income quartile
The assumption going in was that fixing fairness costs accuracy. For income, there's a real trade-off curve, visible here. For home ownership, the opposite happened: calibrating for fairness improved accuracy too, because the default 0.5 threshold was never a good operating point to begin with.
Side-by-side bar charts of the top 10 SHAP features for Lending Club versus Home Credit
The same architecture, unmodified, applied to a second lender's entirely different dataset. Not one feature name overlaps, they don't share columns, but the pattern is identical: both models lean hardest on external credit-bureau-style scores. That's what "the framework generalises" looks like in the data.
Diagram of the EICS Framework: three pillars (Prediction, Explanation, Fairness) plus two EU AI Act compliance layers
The final deliverable: three pillars, prediction, explanation, and a fairness audit, wrapped in two compliance layers matching the EU AI Act's transparency (Article 13) and robustness (Article 15) requirements. Validated on 1.35 million records, replicated on 307,000 more.