Part III — Credit & Wealth

Chapter 18 · Underwriting with Alternative Data

Chapter 17 left the sanction box closed; this chapter opens it. Why the bureau paradox locks first-time borrowers out, how cash-flow lending on Account Aggregator, GST and UPI data replaces collateral with observed reality, what the OCEN and ULI rails are actually building — and the model risks that decide whether a thin-file score widens inclusion or merely automates exclusion.

📖 Chapter 18 of 35 ⏱ ~60 min read 🎓 Session 21 · CILO 2 Core 60% Stable 25% Volatile 15% 📝 Quiz · 30 MCQs

🎯 Learning Objectives

1. The Bureau Paradox

India's credit bureaus — CIBIL first, operating from 2004, then Equifax, Experian and CRIF High Mark under the Credit Information Companies Act, 2005 — are genuinely great at their job: reporting who has borrowed formally and repaid. That is also their limit, and the limit has a name, the bureau paradox: bureaus can only see people who already have formal credit. The first-time borrower — the kirana owner running ₹4 lakh a month through UPI, the gig worker with three income streams, the graduate with no history — is new-to-credit (NTC): invisible to the score exactly when visibility matters most.

The paradox is Chapter 2's information gap, reprised at machine scale. Then, the moneylender saw the borrower's life and the bank saw nothing; now the bureau sees the borrower's past and nothing else. And a score-only lender does something worse than decline the NTC applicant — it declines her automatically, at zero marginal cost, across every lender using the same score, creating the same answer everywhere without anyone deciding it. The NTC gap is why credit depth stalls near half of GDP while payment volumes compound: the rails reach everyone, the score reaches the already-borrowed.

📝
The question this chapter answers: what would a lender need to see to underwrite a person with no credit history — and what new risks walk in when the seeing is done by a model instead of an officer?

2. Three Paradigms: Collateral, Character, Cash Flow

Underwriting is a prediction — will this money come back? — and history offers three ways to ground the prediction. Collateral-based underwriting stops predicting and starts seizing: pledge an asset whose value covers the loss. It is safe for the lender and radically exclusionary — Chapter 2's finding, that the people who most need credit own the least pledgeable property, is its permanent signature. Character-based underwriting predicts from reputation: the bureau score, the references, once the local banker's knowledge of the family. It works beautifully for the already-scored and reproduces the bureau paradox for everyone else.

Cash-flow underwriting predicts from observed flows of money: what comes in, what goes out, how steady, how seasonal, how growing. The insight is old — moneylenders and SHG circles always lent against seen flows — but until the stack existed, formal lenders could not see flows at scale. Now they can: consented bank statements through the Account Aggregator, GST filings for businesses, UPI merchant collections, salary credits. The paradigm shift in one sentence: from "what do you own and what have you borrowed?" to "what does your money actually do?" — and it is the paradigm every rail in the next two sections exists to serve.

📋 Stable content — Reviewed: November 2026 · Next review: May 2027

3. The Signals: AA, GST, UPI

Read each signal with Chapter 10's discipline: what it shows, and — more important — what it does not prove.

SignalWhat it showsWhat it does NOT prove
AA bank statementsReal inflows/outflows over months; regularity; existing EMI burden; seasonalityThat flows are the borrower's own; that past flows persist; undisclosed cash income or obligations
GST filings (GSTN)Declared business turnover, filing regularity, input-output chain — third-party verifiedProfitability (turnover ≠ margin); informal sales below the tax line; a filing rented for the purpose
UPI merchant flowsDaily collections, customer count, growth — the shop's actual heartbeatPersonal vs business mixing; refunds and circular transactions; season inverts
Bureau (for the scored)Repayment history, existing exposure, enquiriesAnything about the new-to-credit file — its blind spot is the point of this chapter
Salary creditsEmployer, income stability, employer healthJob security; gig income's true volatility; second employments

Two disciplines follow. Triangulate: any single signal can be gamed or misread; the kirana owner's UPI collections matter precisely when they match her GST declarations and bank credits — three signals agreeing is worth more than any one of them alone. Convert flows to capacity, not to optimism: a monthly inflow is not a monthly repayment capacity; the hands-on exercise makes the conversion arithmetic explicit. And note what is deliberately absent from the table: device data, contact lists, social profiles — signals some markets once used — because Indian data-protection law and the digital lending rules confine collection to need and consent, a boundary Chapter 31 polices.

⚠ Volatile content — Reviewed: November 2026 · Next review: February 2027

4. The Rails: OCEN and ULI

Two pieces of infrastructure assemble these signals at national scale — and both are younger than this course.

OCEN (Open Credit Enablement Network, pilots from 2022) attacks the integration problem: instead of every digital lender building bilateral pipes to every originator and data source, OCEN provides common APIs so any lender plugs into any originator — credit assembled like Lego, in the embedded settings of Chapter 16's model table. ULI — the Unified Lending Interface — is the bigger sibling: piloted from August 2023 as the "public tech platform for frictionless credit" and formally unveiled by the RBI Governor on 26 August 2024. Built by the RBI Innovation Hub, it is a consent-based, plug-and-play API layer through which a lender draws verified borrower data — Aadhaar e-KYC, PAN, GSTN, bureau, DigiLocker, digitised land records, even satellite crop data and milk-federation flows — in real time, replacing the two-to-four-week appraisal cycle. The Governor named the trinity outright: JAM for identity, UPI for payments, ULI for lending — each solved once, at the infrastructure layer.

The rollout is no longer a pilot. By early 2026, 89 lenders were live — a near-tripling from 36 a year before — across 136+ data services supporting a dozen loan journeys: Kisan Credit Cards, crop and dairy finance, micro-business and gold loans, eMudra, personal, pension, vehicle, housing and MSME lending. Roughly ₹27,000 crore was disbursed through ULI in 2024 across six lakh accounts; individual banks now originate hundreds of crores a month on it; and a full national rollout has been targeted for late 2026. Treat every number in this paragraph with Chapter 10's six-trap discipline — they age quarterly, which is why this section carries a quarterly flag.

Sources: RBI press releases and the RBI@90 unveiling (26 August 2024); DFS review meetings and industry reporting, 2025-26. ULI differs from the AA: the Aggregator is a regulated data pipe from any FIP; ULI is a lender-facing platform with a curated data-service catalogue and loan journeys.

5. Model Risk in Thin-File Scoring

Cash-flow scoring widens the seen world — and automates the judgements made about it. That automation carries risks no disclosure clause of Chapter 17 reaches, because they live inside the sanction box.

Proxy discrimination: a model barred from using community or gender can reconstruct them — pincode proxies neighbourhood, neighbourhood proxies community, and a "risk factor" is born that no one wrote down as bias. Feedback loops: decline the thin-file borrower and she stays thin-file; the model's world confirms itself — the score's blind spot becomes the borrower's ceiling. Drift: a model trained on 2024's flows meets 2026's economy — festive shocks, a new MDR, UPI credit products — and quietly misprices while its accuracy dashboards lag. Gaming: every signal invites its counterfeit — rented GST filings, circular UPI transactions, statement factories. Opacity in decline: an automated rejection with no reason is both cruel and unfixable — the borrower cannot correct a score she cannot see, and the lender may not be able to explain it either.

The governance answer is the RBI's FREE-AI framework (2025) — fairness, responsibility and ethics in the enablement of AI — which makes model risk a board-level duty for regulated entities: documented models, monitored drift, explainable declines, human escalation. Chapter 31 opens it in full; here, note the placement — the same stack that makes inclusion possible at scale makes exclusion possible at scale, and the difference is governance.

⚠
Warning: "The model decided" is never an answer — it is a decision someone made earlier and hid. In every exercise in this course, the model is yours: you own its inputs, its cut-offs and its errors.

6. The Underwriter's Discipline

Close with the working rules that survive every paradigm. Signal ≠ score: data is an input; the conversion — flows to capacity, capacity to amount, amount to price — is a chain of assumptions an underwriter must be able to state aloud. Capacity is a fraction, not a flow: lend against a share of net surplus, never against gross turnover. Volatility is information: two borrowers with the same average inflow and different ranges are not the same risk. Price includes the loss: the rate must carry expected loss plus funding plus cost — Chapter 8's spread arithmetic, now with your own data in it. Decline with a reason: if you cannot explain the no, you have not finished underwriting.

The hands-on applies every rule to one real thin file. And this chapter's engine powers everything Part III builds next: Chapter 19's products — BNPL and credit on UPI — are cash-flow underwriting wearing consumer packaging, with a classification question riding on top.

🛠 Hands-On: The Cash-Flow Underwriting Exercise

This is the session's classroom activity. Meet Meena, who runs a kirana store. She has never borrowed formally (NTC). Her consented signals, six months: UPI merchant collections — ₹4.2, 4.6, 3.9, 5.1, 4.4, 4.8 lakh; GST filings — all six months filed on time, declared turnover averaging ₹4.3 lakh; AA bank statement — credits matching the UPI series, no existing EMIs, rent of ₹15,000 a month. Industry benchmarks put a kirana's net margin near 8 per cent.

Steps

  1. Read the signals. Compute average monthly inflow and the range (max minus min). What does the GST–UPI match tell you? What does it still not prove?
  2. Convert flow to capacity. Estimate monthly net cash (margin on turnover, less rent), then apply a 40 per cent instalment-to-surplus cap. What EMI can Meena carry?
  3. Size the loan. At 24 per cent per annum for 12 months (EMI factor ≈ 0.0946 per rupee of principal), what principal does that EMI support? State the loan as a multiple of monthly turnover and sanity-check it.
  4. Price the risk. With funding at 10 per cent, servicing ₹150 per loan, expected loss at 4 per cent of exposure, does 24 per cent clear Chapter 8's contribution arithmetic? What first-loan size would your model actually start with?
  5. Audit the model. Name three model risks from Section 5 that apply specifically to this file, and one design choice for each that mitigates it.
View Solution / Walkthrough

1. Average inflow = (4.2+4.6+3.9+5.1+4.4+4.8)/6 = ₹4.5 lakh/month; range = 5.1−3.9 = ₹1.2 lakh (roughly ±13 per cent around the mean — steady, mildly seasonal). The GST–UPI match (4.3 vs 4.5, within ~5 per cent) is triangulation: two independent sources agree, so fabrication is less likely and informality is limited. Still unproven: margin (turnover says nothing about profit), cash sales below the tax line either way, and private obligations a statement cannot show.

2. Net cash ≈ 8% × 4.5 lakh = ₹36,000, less rent ₹15,000 → surplus ≈ ₹21,000/month. At the 40 per cent cap: EMI capacity ≈ ₹8,400/month. (The cap exists because the surplus also feeds her household — the 60 per cent is food, school, shocks.)

3. Principal = 8,400 / 0.0946 ≈ ₹88,800 — call it ₹85,000–90,000, roughly one-fifth of monthly turnover. Sanity check: a small multiple of turnover is the natural first loan for an NTC file; ₹5 lakh on a ₹4.5-lakh-turnover shop would be optimism, not underwriting.

4. On ₹88,800: revenue = 24% × 88,800 × ½ ≈ ₹10,700; funding at 10% ≈ ₹4,440; expected loss 4% ≈ ₹3,550; servicing ₹150 → contribution ≈ ₹2,560 — positive on the first loan, unlike Chapter 8's ₹0 example, because cash-flow data shrank priced risk. A prudent model still starts smaller (₹50,000–60,000) and scales on repayment — the repeat-lending logic — with the UPI autopay mandate collecting each EMI at near-zero cost.

5. Strong audits pick three, with mitigations: proxy discrimination — the model must not let pincode stand in for community; mitigated by protected-class blind testing of cut-offs. Feedback loop — a decline keeps Meena NTC forever; mitigated by offering a smaller starter loan instead of binary reject. Gaming — the GST–UPI match deters rented filings, but circular UPI trades could fake collections; mitigated by customer-count and ticket-size dispersion checks. Drift — festive season inflates month five; mitigated by seasonal adjustment before averaging. Full credit names the risk on this file, not in the abstract.

Key Takeaways

1

The bureau paradox: scores see only the already-borrowed. The new-to-credit gap is Chapter 2's information gap at machine scale.

2

Cash-flow underwriting asks "what does your money do?" — AA statements, GST filings, UPI collections — replacing collateral with observed, triangulated flows.

3

ULI (unveiled 26 Aug 2024) is lending's UPI: 89 lenders, 136+ data services, ₹27,000 crore disbursed in 2024 — JAM for identity, UPI for payments, ULI for lending.

4

Thin-file model risk is real: proxy discrimination, feedback loops, drift, gaming, opaque declines — governed now by the FREE-AI framework's board-level duties.

5

Underwriting discipline: signal ≠ score, capacity is a fraction of net surplus not gross turnover, volatility is information, and a decline without a reason is unfinished work.

Test Your Understanding

Choose one best answer per question. A calculator is allowed for the arithmetic items; Meena's file from the hands-on is the data. Aim for 24/30 before moving to Chapter 19.

1. The bureau paradox is that credit bureaus:

2. A borrower with no formal credit history is called:

3. Why is an automated NTC decline worse than an officer's decline? (Hand arithmetic allowed.)

4. India's first credit bureau was:

5. Cash-flow underwriting predicts repayment from:

6. The one-sentence shift cash-flow underwriting makes:

7. Which limitation belongs to AA bank-statement signals?

8. GST-based signals do NOT prove: (Hand arithmetic allowed.)

9. Meena's UPI collections (avg ₹4.5 lakh) match her GST declarations (avg ₹4.3 lakh). This is valuable because:

10. ULI was formally unveiled by the RBI Governor on:

11. The "trinity" of digital public infrastructure named at ULI's launch:

12. As of early 2026, the number of lenders live on ULI was about:

13. Which pair of data sources sits on the ULI platform? (Hand arithmetic allowed.)

14. ULI differs from the Account Aggregator in that ULI is:

15. OCEN's contribution to the credit stack is:

16. A model using pincode as a risk factor may be guilty of:

17. The feedback-loop risk in thin-file models is that declines:

18. Model drift in cash-flow underwriting means the model:

19. Signal gaming in this space includes:

20. The RBI framework making model risk a board-level duty is:

21. The opacity risk in automated declines is that the borrower:

22. Meena's average monthly UPI inflow and range are: (Calculator allowed.)

23. Meena's monthly surplus and 40-per-cent EMI capacity: (Calculator allowed.)

24. At 24 per cent, 12 months (EMI factor 0.0946), the supportable principal is about: (Calculator allowed.)

25. The underwriter's rule on capacity is to lend against:

26. Two borrowers share a ₹4.5 lakh average inflow; one ranges ₹3.9–5.1 lakh, the other ₹1–8 lakh. The correct reading:

27. The mitigation for the feedback-loop risk on Meena's file is:

28. Against circular-transaction gaming of UPI flows, the practical check is:

29. The chapter's first rule of underwriting discipline is:

30. The products built on this underwriting engine, taken up next, are: