A file that opens successfully is not necessarily ready for analysis. A calculation can be correct for the rows in view and still answer the wrong question because the data came from a different account, covers a different period, or contains duplicate records.
I check four things before drawing a conclusion.
1. Does the source match the question?
Confirm that an export came from the intended system, account, and view. A filename or local cache is a clue, not proof of the active source. If several accounts or systems are involved, record the scope of each dataset before combining them.
2. What period does each file cover?
Record the first and last dates, the time zone, and any missing days. Distinguish a day with no events from a day that was never exported. When comparing two reports, start with their overlapping period. Whole-file totals may differ simply because the beginning or end dates do not match.
3. What does one row represent?
Define the grain: perhaps one transaction, one closed position, or one daily total. Do this before counting or averaging. If rows appear duplicated, record counts before and after deduplication and the rule used. Do not remove lookalike rows without checking whether the source can legitimately produce separate events with the same values.
4. Are the fields sufficient for the conclusion?
Check for the identifiers, timestamps, costs, and states needed to answer the question. Without an identifier linking entries and exits, the share of profitable rows is not a position-level win rate. Without complete costs, a net return should not be presented as certain.
Report the evidence boundary
The final analysis should state both what can be calculated and what remains unknown, including the deduplication rule, period used, and missing fields. A pattern in historical data is a question for further testing, not an automatic decision rule.