IDBI Innovate 26 · PS4
DRISHTI
See stress 12 months early — and know what to do
One engine for four loan books: calibrated PD → RG1–10 → Red/Amber/Green →
SMA watch → action playbook → ECL. Decision support only — underwriter decides.
This Space demos India MSME live; other books’ metrics are in Validation.
1 Ingest
GST · AA · bureau · notes · graph → 2 Canonical schema
IDBI-swappable seam → 3 Segment models
LightGBM · TabPFN · survival → 4 Interpretation
PD · grade · SMA · action · ECL → 5 Explain & act
SHAP · RBI EWS · cockpit / API
GST · AA · bureau · notes · graph → 2 Canonical schema
IDBI-swappable seam → 3 Segment models
LightGBM · TabPFN · survival → 4 Interpretation
PD · grade · SMA · action · ECL → 5 Explain & act
SHAP · RBI EWS · cockpit / API
India MSME holdout: calibrated AUC — · top-10% of book captures 57% of eventual stress (vs ~16–22% status-quo). Accuracy ≥90% at documented operating points — see Validation tab.
Step 1 — choose an account
Sample India MSME account
Legend:
RAG = Red / Amber / Green stress bucket ·
RG = risk grade (RG1 safest → RG10 riskiest) ·
PD = calibrated 12-month probability of stress ·
Changing the account updates the verdict instantly (no Score button).
Step 2 — read the verdict
Recommended action: Enhanced monitoring; request updated stock/GST statements
Calibrated 12-month PD
21.6%
≈ 216 in 1,000 similar accounts
Risk grade · RAG
RG6
Amber
· elevated
SMA watch
SMA-0 watch
Expected stress onset ≈ month 7
Expected credit loss
₹0.01 Cr
Owner · Credit analyst · Monthly review
Model vs policy: aligned (RAG grade and RBI EWS point the same way).
Recourse if borrower improves: Cure EMI bounces (standing-instruction move + buffer) → PD 21.6% → 1.2% · ECL saves ₹0.01 Cr
Signal families present: base, cashflow, notes, graph (4/4 families present)
How to read this: calibrated PD → risk grade (RG1–RG10) → Red/Amber/Green → SMA watch → action playbook → ECL. SHAP = model drivers; EWS = RBI-style policy rules. Underwriter decides.
Why the model says this (SHAP)
REPAYMENT_BEHAVIOUR | Repayment behaviour (EMI / utilisation) | increases risk | -0.0349 |
Policy flags (RBI EWS rules)
EWS01 | Broken repayment promise recorded by the officer | medium |
India MSME data plane (what IDBI already has access to):
GST compliance · Account Aggregator / UPI cash-flow · bureau · officer notes · supplier graph.
Structured-only ≈ today’s 16–22% world; GST/AA + notes are the lift staircase in the deck.
| Field | Value |
|---|---|
| loan_id | IN112628 |
| ticket_size | 857175.5994672261 |
| emi_bounce_6m | 2 |
| gst_filing_delay_days | 0.5020344305510454 |
| cashflow_volatility | 0.8283931132919835 |
| current_ratio | 2.0882432707744814 |
| segment | msme_india |
Explore sensitivity. For full SHAP + EWS story, use a sample account above.
50000 50000000
0 12
0 120
0 1.5
0.2 3
Why not accuracy alone?
At a ~10% default base rate, always predicting “no stress” looks ~90% accurate
and catches zero stressed accounts. Meet ≥90% accuracy at a documented
operating point, then report AUC and capture@top-10% — how many real defaulters
sit in the riskiest 10% of the book (typically 3–5× the 16–22% status quo).
Which book’s metrics?
Model health — MSME — India (synthetic, GST/AA/notes/graph)
| Model | AUC (cal.) | PR-AUC | Capture@top-10% |
|---|---|---|---|
| lightgbm_raw | 0.8562 | 0.3393 | 58.1% |
| lightgbm_calibrated | 0.8555 | 0.3202 | 57.3% |
Top PSI (train vs test):
interest_spread_bps: 0.0018current_ratio: 0.0012gst_filing_delay_days: 0.0011tenure_months: 0.0010ticket_size: 0.0009
Operating points — max accuracy 96.3% (meets the ≥90% requirement)
operating_point threshold accuracy recall_capture flag_rate
default (PD >= 50%) 0.5000 0.9634 0.1186 0.0069
F1-optimal 0.2279 0.9549 0.3620 0.0344
accuracy-optimal 0.4516 0.9628 0.1820 0.0125
top-10% review capacity 0.1140 0.8974 0.6012 0.1105
India MSME samples are synthetic (SIDBI–TransUnion priors). Decision support only.
Full Streamlit cockpit (portfolio ECL, scenarios, contagion, governance), four segments,
and FastAPI live in the GitHub repo.