Content authored by practitioners with experience at Amazon India, Intuit QuickBooks, and the Tata Group. Meet the team →
Lenders adopt bank-statement-based MSME scoring and synthetic financials on the strength of vendor demos, without checking whether the tool's view at application would have ranked their own borrowers' later repayment correctly.
Replay the tool blind on around 200 closed MSME loans using statements held at application, group results into tool bands, compare 90-day delinquency by band, and note segments the tool cannot assess.
Closed-loan sample with application-date statements, outcomes and the decision taken then; outcome fields hidden from the tool; a bias log covering survivorship and approved-only data; owners for each step.
A band-by-delinquency table, a list of borrower types where the tool cannot form a view, analyst override patterns and a clear statement of how far the credit head can rely on the tool.
Before you rely on synthetic financials or bank-statement signals for MSME credit, replay the tool on loans you have already closed. Give it the statements you held at application date, keep the outcomes hidden, and compare its view with what actually happened. Judge whether it ranks risk in the right order across the book; no tool will be right on every single loan.
For background on how these figures are built, see synthetic financial statements for MSMEs and MSME credit without audited financials.
How do you build the back-test?
Pull closed MSME loans, typically disbursed 18 to 36 months ago so outcomes have had time to show. For each loan you need:
- The statements held at application, exactly as received, not a later download.
- The outcome: repaid on time, 30 days late, 90 days past due, restructured or written off.
- The decision taken then: amount, tenor and any conditions.
- Product and segment: for example term loans and working-capital lines, and borrower size under the MSME classification revised from 1 April 2025, under which a micro enterprise has investment up to ₹2.5 crore and turnover up to ₹10 crore, a small enterprise up to ₹25 crore and ₹100 crore, and a medium enterprise up to ₹125 crore and ₹500 crore.
Keep the tool blind to outcomes. Send only the statements and application data; hold outcomes in a separate sheet that only your risk team sees until the read-out.
| Step | Data needed | Bias risk | Safeguard | Owner |
|---|---|---|---|---|
| Select loans | Closed MSME loans, 18 to 36 months old | Only approved loans are in the sample | State plainly that declined applicants are not tested | Risk |
| Gather statements | Application-date statements | Later statements leak outcome information | Use only files dated before disbursal | Credit operations |
| Hide outcomes | Repayment history | Vendor tunes to known outcomes | Outcomes held by risk until read-out | Risk |
| Replay | Statements and application data | Survivorship: loans sold or closed early drop out | Record excluded loans and why | Vendor and credit analyst |
| Read-out | Tool bands and outcomes | Small counts of bad loans | Report counts, not just percentages | Credit head |
Remember throughout that synthetic financials are approximations built from bank flows. They are not audited statements and do not replace them.
What do you read from it?
Three things:
- Ranking quality. Do weaker tool bands line up with higher later delinquency? Look at the order across bands, not a single loan.
- Blind spots. Which borrower types the tool cannot assess well: cash-heavy traders whose sales never touch the bank, and proprietors who run business through a personal account. The personal vs business separation article explains why mixed accounts are hard.
- Overrides. Where your analysts approved against what the tool would have suggested, and how those loans performed.
Worked example: 200 closed loans
Illustrative figures. A lender replays 200 closed MSME loans across two products and groups the tool’s output into three bands.
| Tool band | Loans | 90 days past due | Rate |
|---|---|---|---|
| Strong | 80 | 3 | 3.75% |
| Moderate | 70 | 7 | 10% |
| Weak | 50 | 10 | 20% |
| Total | 200 | 20 | 10% |
Delinquency rises steadily from strong to weak, so the tool ranks this approved book in the right order. What the credit head can safely conclude: the tool’s bands are useful for pricing, limits and review intensity on borrowers like these. What the credit head cannot conclude: that the tool would have done well on declined applicants, or that a strong band guarantees repayment, since 3 of 80 strong-band loans still went 90 days past due.
Then split the result by segment. If 12 of the 50 weak-band loans are cash-heavy traders whose statements show little business activity, the tool may be saying “cannot tell” rather than “high risk” for that group, and the policy should route them to other evidence such as GST returns.
How should the result change your policy?
Write the conclusion as a reliance statement: which products and segments the tool may be used for, which bands trigger manual review, and which borrower types need other evidence. Repeat the back-test once a year on newly closed loans, because borrower mix and bank formats change.
To check the tool itself before the back-test, the same buyer test applies:
| Test | What to give the vendor | What a pass looks like |
|---|---|---|
| Synthetic financials | Statements for 5 borrowers with audited accounts | Figures labelled as approximations; differences from audited accounts explained |
| Account mixing | A proprietor’s personal account used for business | Personal and business flows separated, with the rows shown |
| Blind replay | Application-date statements only | Output produced without outcome data |
How Terra Insight fits
TransactIQ builds synthetic P&L, balance sheet and cash flow from bank statements for MSME underwriting, alongside 40+ engineered credit signals, and runs as a managed service on Terra Insight’s AWS Mumbai infrastructure. These synthetic financials are approximations to read with other evidence, not audited statements. Evaluating teams can send a representative sample for a fit benchmark; see MSME lending.