Time Series Pipelines: Rolling Windows & Resampling

[!NOTE] Time-series telemetry from IoT sensors or trading platforms arrives with timestamps in varying timezones, out-of-order records, and missing sample intervals. Pandas provides native C-optimized time-series engines.

1. Key Time Series Operations

• resample('5min').mean(): Downsamples high-frequency data into discrete 5-minute buckets.
• rolling(window='1h').sum(): Computes a continuous rolling metric over a 1-hour temporal window.
• tz_localize('UTC').tz_convert('America/New_York'): Handles daylight saving time transitions without arithmetic bugs.

Time-Series Resampling, Rolling Windows & Timezone Safety

• Strict UTC Storage Rule: Always localize naive timestamps immediately at ingestion (tz_localize("UTC", ambiguous="infer") or utc=True), convert with .tz_convert(...) only for user presentation, and NEVER mix timezone-aware with timezone-naive columns in joins!
• Rolling Without Leakage: For ML fraud/forecast features, verify whether closed="left" is needed on time-based rolling windows (df.rolling("7D", closed="left")) so current-row outcome doesn't leak into its own lookback feature!
• Resampling vs Grouper: df.resample("1H", on="event_ts").sum() creates continuous time buckets safely—filling empty intervals with 0 rather than silently skipping quiet hours.