An interest rate risk spreadsheet reports that a 100 basis point rise adds Rs 12 crore to earnings. Review finds a gap bucket with the wrong sign and a rate hard-coded from last year; corrected, the answer is a Rs 9 crore loss. How are such errors found, and what controls do end-user models need?

The sheet reported plus Rs 12 crore because a flipped sign turned a Rs 7.5 crore cost into a gain and a stale base rate doubled the shock; corrected, a 100 bp rise costs about Rs 9 crore. Errors like these are caught by sense checks against the gap profile, independent rebuilds and formula audits. Any spreadsheet feeding a decision needs an owner, a tier, locked inputs linked to source, reconciliation and independent review, the same controls as a system model.

Separate them, because a reviewer who fixes one error and sees a plausible number may stop. The flipped bucket alone turns a Rs 7.5 crore cost into a Rs 7.5 crore gain, a Rs 15 crore swing at a 1% shock; the stale base rate alone doubles every bucket's effect. Together they produce plus Rs 12 crore. Fix only the rate and the sheet says plus Rs 6 crore; fix only the sign and it says minus Rs 18 crore, twice the truth. Only both give minus Rs 9 crore.

Imagine a household budget that says your savings rise when the rent goes up. You would not need to check the formulas to know something is wrong; you would know from what the numbers mean. The fastest check is the sign: Yashvel's own gap report shows more liabilities than assets repricing within the year, so rising rates must cost it money, and a positive answer is a contradiction. Then three more. Back out the shock: dividing the output by the weighted gaps shows 2%, not the 1% the scenario claims. Reconcile each bucket's gap to the system extract it came from. And audit formulas for typed constants: any number inside a formula or a hard-coded input with no source is a finding. A quarter-on-quarter movement check would also have flagged a sensitivity that flipped sign with no change in the balance sheet.

The same as system models, scaled to their use. A spreadsheet that shapes ALCO decisions is a tier 1 model in everything but name, and the medium does not lower the bar. Put it in the model inventory with an owner and a tier. Keep inputs on one sheet, each linked to its source or labelled with source and date, and lock the calculation cells. Build in automatic checks: the effective shock equals the scenario, gaps reconcile to the source report, and the sign agrees with the gap profile. Use version control so the ALCO can see what changed. And require independent review before first use and after any change, which is exactly what found these errors, a quarter late.

Close with the consequence: the ALCO spent a quarter believing rising rates helped, and may have chosen not to hedge on that basis. The cost of the error is the decision it supported, not the arithmetic.