Prediction timing and analytical evidence6 min read
Keep the limits attached to the result
Historical records contain knowledge accumulated after a decision. A prototype must be tested using only what could have been known when its proposed answer was needed.
A complete row count can coexist with incomplete observations. Reporting should preserve the difference between an observed absence and a history the source never supplied.
03
Analytical access and small-group disclosure6 min read
Removing names does not settle the suitability of a dashboard. Its filters, audience and surrounding knowledge can make a small group unexpectedly revealing.
A leadership conviction
Before asking what a number predicts, I ask when it became available, which population it describes and what absence means in its source. A polished result can be wrong for the decision even when every calculation is correct. My responsibility is to preserve those distinctions and make them understandable to the people who must act.
Bank branches
86
Three central Indian states
Data colleagues
32
Four managers; March 2026
Annual function expense
₹8.4 Cr
Data office; no credit or treasury delegation
I lead a data office that keeps the timing, population and limits of evidence attached to the bank's analytical decisions.
I am chief data officer of Varneli Cooperative Bank in Bhopal. Our member-owned institution serves households and small enterprises through 86 branches in three states. At March 2026 it held ₹6,800 crore of assets and employed 1,250 people. I lead 32 data colleagues through four managers, with an ₹8.4 crore annual budget. I report to the CEO and present data issues to the board as an executive. Credit, risk validation and financial reporting decisions remain with their authorised owners.
I began in survey analysis in 1997, where an unanswered question had to remain different from a recorded answer of zero. Agricultural-service reporting then taught me how seasons and changing territories can alter the meaning of a trend. I joined Varneli in 2008, moved from management analysis into data stewardship in 2014 and became CDO in 2021. The work still begins with questions learnt in those earlier roles: who was observed, when, under which definition and with what missing evidence?
The cases here involve a credit-model prototype that used future knowledge, a growth comparison distorted by changed branch territories, incomplete historical records rendered as zero and a dashboard whose small groups made individuals too easy to infer. Each required giving up an attractive output or delaying a useful tool. I want readers to see how a data leader can make a decision more credible without claiming to own every business judgement the analysis supports.
A strong prototype score did not survive its intended use
A prototype trained on 12,000 completed loan histories used fields created after the proposed decision point. Its reported random-split AUROC was 0.91. Removing those fields and evaluating on a later time period produced 0.71. We withheld release; the change combined a feature correction and a different validation design, not an isolated estimate of one field's effect.
A boundary transfer was not organic growth
A regional dashboard showed 43% account growth after twelve service-area codes moved between nine branches. We rebuilt the comparison on a consistent geographic basis, yielding 8% for the defined population. The original and restated views remained distinguishable; no customer balance or branch record was changed to improve the comparison.
Missing history remained missing
Among 58,200 migrated analytical records, 7,300 lacked the earlier-period observation needed for a trend. A reporting default had displayed those gaps as zero. We replaced the default with an explicit missing-history state and restricted the comparison to supported observations, giving up the appearance of complete coverage.
Current focus
I am strengthening the way analysts describe the population and time available to a decision. The team is reviewing historical coverage, model inputs and small-group reports before wider use. The aim is to make limitations visible at the point where someone relies on an output, while reducing unnecessary circulation of customer-level material.
Conversations I welcome
I would consider a first non-executive role in a regional service or business-information company that wants stronger oversight of analytical evidence. I have provisionally set aside eighteen days annually, subject to employer approval and conflicts review. I would contribute data-governance judgement while learning the wider business, without offering my executive access to bank records or pretending a data title establishes board eligibility.
Beyond the page
The record behind the perspective.
Formation
Bachelor of Science in StatisticsSarnethi Statistical College · 1997
I have worked with Leela as CEO since 2021. She can explain why a result is not ready for the decision we want to make and still give us a practical next step. That is especially valuable when the original presentation was more reassuring than the evidence beneath it.
Nirmal VekshaniCEO, Varneli Cooperative Bank; executive colleague since 2021The full record ↗
Contact
I would consider a first non-executive role in a regional service or business-information company that wants stronger oversight of analytical evidence. I have provisionally set aside eighteen days annually, subject to employer approval and conflicts review. I would contribute data-governance judgement while learning the wider business, without offering my executive access to bank records or pretending a data title establishes board eligibility.
