Chief Data Officer / India
The number changed meaning before it changed value
- Data colleagues
- 18
- Three leads across the CDO remit
- Footwear shops
- 26
- A Srinagar-headquartered retail enterprise
- Annual data-function expense
- ₹4.3 crore
- ₹1.2 crore programme authority stated separately
Decisions & consequences
01Correcting the average-order label on a daily-average calculation
In 2020 I approved a summary that took the simple mean of daily basket averages and labelled it as the average order value for the combined period. A synthetic two-day check exposed the mismatch. Day one had ten orders averaging ₹2,000, and day two had ninety averaging ₹1,000. The simple mean of the two daily averages was ₹1,500. But the underlying amounts were ₹20,000 and ₹90,000, total ₹1.1 lakh across one hundred orders, giving ₹1,100 per order.
The original formula was not meaningless: it described an equally weighted average of the two days’ averages. The error was presenting it as though each order had equal weight in the combined result. I had approved that interpretation and needed to own the correction. We withdrew the affected comparison from the next trading recommendation and asked authorised owners to review where the earlier summary had been used. The synthetic amounts were test evidence, not a claim that ₹40,000 of real sales had appeared or disappeared.
I approved ₹27 lakh over eight weeks to review aggregation definitions, retain underlying totals and counts, and test the relevant reporting paths. The review identified thirty-two uses of the measure. Twenty-five were corrected and accepted under their agreed definitions; seven remained held for unresolved lineage or scope questions. We did not present the twenty-five as proof that every report in the enterprise was accurate. Finance retained its financial-reporting responsibility, and commercial owners assessed whether earlier decisions needed reconsideration.
The repair displaced other analytical work and required me to explain a mistake in a report I had endorsed. It did not create a measured revenue improvement. The durable change was a clearer connection between the question, the weighting and the label. A familiar dashboard heading could no longer be treated as sufficient evidence that a new aggregation answered the same question. The account retained the unresolved seven uses so a partially completed repair did not become another overstated summary of the whole reporting estate.
Agreeing the clock before comparing daily trading results
A 2023 review found that two trading summaries assigned late-evening events to dates on different bases. One grouped the stored UTC timestamp directly, while the other derived the local calendar date used by the retail team. An event at 18:45 UTC on 15 June 2023 corresponded to 00:15 on 16 June in India. Neither representation changed the event itself, but the date label could put it into a different daily total. An agreed trading cutoff could introduce a further distinction from the local calendar date.
The team initially proposed adjusting the display label alone. I asked technology, operations and finance to agree the meaning of the report first. We retained the original timestamp and made the conversion and reporting boundary explicit. The financial owner retained authority over the relevant reporting policy; the data programme did not silently move statutory accounting periods. We used maintained timezone definitions and boundary tests rather than assume that the server’s own setting supplied the business convention.
The ₹41 lakh, ten-week programme reviewed 120 defined boundary examples across the affected flows. Ninety-four met the agreed rules at the review point and twenty-six remained unresolved because source timing or cutoff ownership still needed clarification. Those were test examples, not 120 customers or a count of incorrect financial statements. The unresolved flows stayed outside a claim that the daily account had been reconciled. The programme also required an explanation for commercial readers so they could distinguish a changed definition from a change in actual trading.
We accepted slower publication of affected comparisons while the basis was settled. I did not report an invented cash recovery or claim that timezone conversion alone made the trading account complete. The result was an agreed definition for the accepted flows and visible unfinished work for the rest. My lesson was that a date can look self-explanatory while carrying several choices. Those choices belong with the number before a manager compares one day with another or treats a boundary difference as a business movement.
Leaving the unrecorded return reasons visibly unrecorded
In 2025 a review contained one thousand return records. Six hundred and ten had a recorded reason and 390 did not. Among the 610 recorded reasons,183 were coded as a size issue:30% of the recorded-reason subset, or 18.3% of all one thousand reviewed returns. Neither percentage established the size-issue share among the 390 records with no reason. A proposed chart distributed the unknown group across observed categories to create a complete-looking profile.
I rejected that presentation because it supplied an assumption the records did not establish. We retained the known reasons and a separate unknown group, with both denominators explained. The service team wanted a useful basis for improving the customer process, and I supported that objective without turning missing data into a confident explanation. Privacy and service owners reviewed whether the collection process could be improved appropriately; the data team did not impose a compulsory question merely to make the next chart easier to complete.
We allocated ₹34 lakh over twelve weeks to examine the capture route and clarify the reporting definitions. A selected review of forty service interactions helped identify where a reason could be omitted or recorded inconsistently. Those observations were diagnostic, not a representative explanation of all 390 missing records. The work did not retrospectively invent answers for customers who had not supplied them. Product and buying colleagues retained their decisions about range and fit, and no medical or universal product-suitability conclusion was drawn from the coded reason.
The immediate result was a less complete-looking but more faithful account. Management could see what the records supported and what required another source of evidence. I did not claim that the programme reduced returns or proved a particular product defect. A blank field can arise for several reasons, and the absence of an answer should remain an analytical limit until appropriate evidence changes it. Keeping that limit visible helped the retailer ask a better service question instead of treating a convenient distribution as the customer’s own explanation.