AI Credit Scoring with ML: Thin-File Coverage & Fairness

Traditional scoring (FICO, credit bureaus) evaluates only one in five borrowers with thin credit files — the rest are rejected or given high rates. According to FICO, thin‑file coverage rarely exceeds 30%. ML models with alternative data capture nonlinear signals: regular deposit top‑ups, stable service subscriptions. We have extensive experience building such models in production — delivering a Gini gain of 15–25 percentage points over regression. In our practice, ML scoring reduces missed defaults by 2–3 times compared to traditional methods, and risk loss savings can reach 25% of the overdue portfolio. The pilot project typically pays for itself within the first year through reduced losses.

Which data makes the model more accurate?

Traditional credit data — payment history, limit utilization, inquiries. Necessary but insufficient for thin files.

Alternative sources:

• Open Banking transactions (income stability, spending patterns, savings).
• Phone metadata (top‑up regularity, carrier).
• Browser behavior (with consent).
• Verified employment.

Comparison of approaches:

ML scoring with alternative data does more than improve metrics — it gives the business access to new client segments previously excluded. Specifically, ML scoring covers 3 times more thin‑file borrowers than traditional scoring.

How we build a production‑ready model

We use LightGBM with time‑based split (TimeSeriesSplit), isotonic regression calibration, and fairness auditing. We also fine‑tune hyperparameters with Optuna to achieve maximum Gini. Example pipeline:

import lightgbm as lgb
from sklearn.model_selection import TimeSeriesSplit
from sklearn.calibration import CalibratedClassifierCV

credit_features = ['dpd_30d','dpd_60d','dpd_90d','utilization','num_accounts']
transaction_features = ['avg_income_3m','income_stability','gambling_ratio']
all_features = credit_features + transaction_features

model = lgb.LGBMClassifier(n_estimators=500, learning_rate=0.01, num_leaves=63,
                           min_child_samples=50, colsample_bytree=0.7,
                           class_weight='balanced', random_state=42)
tscv = TimeSeriesSplit(n_splits=5)
scores = cross_val_score(model, X[all_features], y, cv=tscv, scoring='roc_auc')
calibrated_model = CalibratedClassifierCV(model, method='isotonic', cv=3)

After calibration, the model outputs correct default probabilities, critical for credit decisions.

Explainability: why the refusal?

Each rejection is accompanied by a SHAP decomposition — top‑3 negative factors and their numerical contribution.

import shap
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_single)
negative = sorted(zip(features, shap_values[1]), key=lambda x: x[1])[:3]
# Output: high limit utilization, short tenure, 24m delinquencies

This is a mandatory requirement for regulators (ECOA, GDPR). We guarantee compliance with all local rules.

Fairness: how to avoid discrimination?

ML models can replicate historical biases. Mandatory steps:

• Disparate Impact: approval rate across groups must not differ by >20%
• Equalized Odds: TPR/FPR equal for all groups
• Proxy detection: address must not be a proxy for ethnicity

We use adversarial debiasing and reweighing — part of our expertise. We also provide a fairness audit report that can be submitted to regulators.

Monitoring: when to retrain?

After deployment we track PSI (Population Stability Index). If PSI > 0.25 — retrain. We also perform vintage analysis: compare defaults of cohorts by origination month. During crises we dynamically calibrate the baseline default rate.

Step‑by‑step implementation of ML scoring

1. Data audit and feature engineering (2–4 weeks). Assess available credit and alternative sources.
2. Baseline model on historical data (2–3 weeks). Validate Gini improvement.
3. Hyperparameter optimization and fine‑tuning (1–2 weeks). Use Optuna.
4. Probability calibration and fairness audit (1 week).
5. Integration into IT infrastructure (API or batch scoring) and deployment (3–6 weeks).
6. Monitoring setup: PSI, vintage, fairness (1–2 weeks).
7. A/B test and final tuning (4–8 weeks).

What our ML scoring service includes

• Data audit and feature engineering (including alternative sources)
• Baseline and production‑ready model development
• Model integration into your IT infrastructure (API, batch scoring)
• Monitoring dashboards: PSI, vintage, fairness
• Full documentation and team training (2–3 day workshops)
• Support during A/B test and first month after deployment

Timelines and cost

Timeline: 4 to 8 months depending on data volume and regulatory review requirements. A 2–3 month pilot quickly validates hypotheses. Risk‑related savings from ML scoring typically pay back the project within the first year. Contact us for a data audit and precise implementation cost estimate.

Why start with a pilot?

A 2–3 month pilot provides:

• Gini improvement validation on your data
• Assessment of alternative data impact
• Model prototype for internal presentation
• Understanding of full implementation effort

Get a consultation — we will assess the potential of ML scoring on your data. Contact us to launch a pilot project.