10 · Usage & ongoing use · Inside the definition
Yesterday’s performance dropped a slice because someone switched to payment time
The data analyst explains movement on an already frozen definition, and gives the next action. Changing the definition is not analysis. It is another table.
The data analyst works inside a frozen definition
The data architect freezes gross margin, order amount, and time fields. The data analyst does not invent a parallel algorithm. Once the definition is frozen, what he does is compare, drill, and attribute: why this number is higher or lower than yesterday, which documents it can land on, whether the next action is check these three documents, ask this store, or look at this status. A conclusion that cannot drill to a document does not enter the monthly meeting. Analysis that can name a theme in the meeting but cannot match a document number sends people back to Excel.
When he finds the definition is not enough, he submits a change to the data architect and notes from which day it takes effect. Numbers before the effective date are not quietly recalculated on the new definition. A recalculation can be a separate comparison version, but the two versions must have different names. He keeps the old name in the archive so minutes cannot be matched to the wrong one.
An anomaly is a document, not a feeling
A store suddenly at zero: he first sees whether there are no orders, reporting interrupted, or a filter excluded the store. The three treatments differ. A filter error changes the report. It is not going to criticise the store. If zero orders is real, it goes to customer success or product operations, to see whether the store closed or went back to WeChat. The data analyst does not do the visit for them, but he must give the time range and the document-number gap clearly.
The gap caused by payment time
The day after a campaign, sales said performance was short. The data analyst found a temporary report using payment time. The official definition is order time. Orders paid across midnight were counted on the next day. He did not change the official report to payment time to please this campaign. In the note he listed the cross-midnight document numbers and amounts, and asked the data architect to confirm whether campaign review needed a separate payment definition. Until confirmed, the monthly meeting still used the original definition. When two sets of numbers sit side by side, the names differ. Afterwards the business chose payment time for review and order time for appraisal. Both names went into the metric note. The quietly substituted temporary table was deactivated.
The value of this role is stopping the argument on documents. If it will not stop, analysis is not finished. The data analyst writes the definition of every temporary report in the header. A file with no definition in the header does not go out. If someone forwards it with the header deleted, the next version gets a name that cannot be deleted by mistake, for example order amount (order time). “Performance” with no time in the name, he no longer uses. Differences across midnight, across warehouses, and across statuses each get a document-number list. Mix them and you only get a total that looks even. Totals even but document numbers shifted, and warehouse and finance will each think the other is wrong. On every period-over-period anomaly he first asks three filters: time, store, status. If one of the three changed, it is not a business movement. Ask that, then drill document numbers. Turning an anomaly straight into an operating conclusion sends the meeting the wrong way. The cost of the wrong way is not an ugly report. It is the warehouse being told to restock on the wrong date.
He writes the conclusion as a next action that can be verified: check these three documents, ask this store, look at this status. Write only “suggest attention,” and next week’s meeting nobody remembers what to attend to. Analysis ends on an action, not on a chart. He archives the old definition of a same-named metric, and forbids new reports from citing the old name. If the old name stays usable, three months later there will again be two performances that cannot be matched. A forbid on citing is more effective than a reminder to be careful, because the reporting tool will not remember the argument from the meeting.
Changing the definition is not analysis. It is another table. When the definition is not enough, submit it to the data architect with an effective date. Numbers before that date are not quietly recalculated on the new definition. A comparison can be a separate version. The two versions must have different names. A conclusion that cannot drill to a document does not enter the monthly meeting.