MSME Financial Health Card

New-to-Credit and New-to-Bank MSMEs often lack bureau history, audited financials, or collateral records. This self-directed prototype tests whether consent-oriented operating evidence can support an inspectable pre-screening decision without pretending that absence is repayment performance.

The design constraint is practical: use evidence a small firm can plausibly produce, keep structural missingness explicit, and make each downstream judgment reviewable.

The typed contract accepts 33 raw inputs across GST filing and turnover, UPI and banking flows, EPFO and payroll regularity, invoices, and firmographics. Label-block columns stay behind an allowlist, while derived ratios are calculated only after validation.

For the 32,556 synthetic firms without an existing loan, repayment history remains structurally missing. The transparent rubric redistributes unavailable repayment weight across evidence that exists rather than mean-imputing a fictional credit record.

Six monotone pillar rubrics own the Financial Health Score. A statistical model separately estimates the generated probability-of-default target and eligibility; eight named consistency rules surface conflicting GST, bank, UPI, payroll, and invoice evidence.

A versioned policy then owns risk band, knock-outs, limit, tenor, and indicative pricing. The separation keeps a reviewer from mistaking a transparent score, a statistical estimate, a deterministic flag, and a policy choice for one opaque verdict.

For illustrative synthetic record MSME0000042, six pillar contributions add up to 77.6 points. The separate model estimate is a 6.9% generated probability of default; the policy returns eligible, Medium risk, ₹273,000, 24 months, and 13.5%.

Named rubric reason codes explain the score; SHAP names effects on the estimated generated probability of default; consistency rules report broken invariants; and policy explains the recommendation.

The model is evaluated on a fixed 10,000-row synthetic holdout after training on 40,000 separate rows. Eligibility discrimination, generated-PD error, Financial Health Score fit, and every pillar rubric are read against configured gates.

Ten equal-frequency bins compare predictions with the generated PD target—not observed defaults. Customer segment, industry, location, and business type slices retain their row counts, PD MAE, and signed bias so stability can be inspected without claiming fairness certification.

The simulator lets a reviewer adjust explicit, controllable levers—such as GST filing timeliness or overdraft use—then compare score, generated-PD risk, band, limit, and pricing with the baseline.

Only the scenario is re-scored. The stored synthetic MSME record remains unchanged, so sensitivity can be inspected without presenting the simulated recommendation as an offer.

This self-directed synthetic-data prototype demonstrates a typed alternative-data contract, leakage-aware model evaluation, exact score arithmetic, calibrated generated-target behavior, versioned recommendations, and reviewable failure signals. It does not demonstrate production deployment or real borrower outcomes.

The highest-value next experiment is prospective validation on consented operating data joined to observed repayment performance: freeze the feature, rubric, and policy versions; measure calibration and error by cohort over time; then retrain or reject the generated-PD mapping before any real credit use.