Dates and missing values change grouped results

Understand the likely causes

• Invalid dates may become NaT, naive local clocks may be labeled UTC, resample boundaries may differ, or missing observations/group keys may leave the denominator. Text missing-token rules can also alter identifiers or measures at ingestion.

Preserve raw tokens and the intended reporting clock

Record the source snapshot, raw time tokens, parser options, original missing markers and exact library versions. State whether the report uses UTC days or a named local business calendar. A date-looking string does not identify its timezone or guarantee one strict interpretation. dayfirst is a preference rather than a strict format validator. utc=True localizes naive inputs as UTC, so it cannot repair an unknown local zone automatically. Preserve source context before normalizing values; otherwise a plausible normalized column can hide the original ambiguity and make historical reconciliation impossible.

Count new parse failures and changed denominators

Compare known missing timestamps with newly introduced NaT from coercion. Track valid source rows, rejected rows and rows entering each group; do not let an invalid parse silently disappear. For measures, distinguish row count from nonmissing count and an observed zero from an unknown value. Missing group keys can be excluded by groupby policy, and all-missing sums need their own min_count interpretation. A lower daily total might reflect changed coverage rather than real inventory movement. Keep rejected identities and reasons so a change can be explained without reconstructing the source from a final aggregate.

Compare the same instant across the chosen calendar

Timezone conversion can move an instant into another displayed calendar date without changing the instant. The local offset example crosses from October first to September thirtieth in UTC. Extracting date before versus after conversion therefore defines different populations. Localization near daylight-saving transitions adds ambiguous/nonexistent-time choices that need a documented policy. For resampling, inspect frequency, closed, label and any origin/offset relevant to the chosen bins. An observation exactly on a boundary must have an expected period. This model checks one standard-library conversion only and does not certify pandas timezone or resampling behavior.

from datetime import datetime, timezone

local = datetime.fromisoformat('2026-10-01T01:00:00+02:00')
utc = local.astimezone(timezone.utc)
assert local.date().isoformat() == '2026-10-01'
assert utc.date().isoformat() == '2026-09-30'
print(local.date().isoformat(), utc.date().isoformat())

2026-10-01 2026-09-30

Accept the report through identity and temporal coverage

Build a small fixture with a valid zero, a missing measure, an invalid date, an exact bin edge and a local instant near a day boundary. Define expected accepted/rejected identities, period assignments and observed denominators independently. Compare the structured result before formatting dates or filling blanks for presentation. This guide runs no pandas parser or resampler, installs no timezone backend and reads no real file. After an authorized runtime correction, preserve the original source, rule version and explicit accepted gaps. A smooth complete-looking report is not acceptance unless its population and calendar contract are verified.

Keep it from recurring

• Version parser, missing-token, timezone and binning rules with the analytical source contract. Retain exact-edge and all-missing fixtures across library upgrades, and monitor parse-failure coverage separately from report formatting.