Time Series: Dates, Resampling and Rolling Windows

What you'll build

Daily and weekly revenue from individual orders, with days that had no orders filled in as zero, a 3-day rolling average to smooth noise, and day-over-day change: the core of any metrics chart.

The idea

• Use a DatetimeIndex (set_index("date")) for time-based operations.
• resample("D"), "W", "MS" groups by calendar periods; empty periods appear (fill them with asfreq/fillna(0)).
• rolling(window) computes moving statistics; pct_change() and diff() compute changes.
• .dt gives date parts; tz_localize/tz_convert handle time zones.

The code

from io import StringIO
import pandas as pd

csv = StringIO("""order_id,date,customer,country,product,quantity,unit_price
1001,2026-01-03,ana,DE,mug,2,8.50
1002,2026-01-03,ben,US,teapot,1,24.00
1003,2026-01-04,chen,DE,kettle,1,31.90
1004,2026-01-05,ana,DE,teapot,1,24.00
1005,2026-01-07,dara,FR,mug,4,8.50
1006,2026-01-07,ben,US,mug,1,8.50
""")
orders = pd.read_csv(csv, parse_dates=["date"])

orders["total"] = orders["quantity"] * orders["unit_price"]
daily = orders.set_index("date")["total"].resample("D").sum()      # days without orders become 0
report = pd.DataFrame({
    "revenue": daily,
    "rolling_3d": daily.rolling(3, min_periods=1).mean().round(2),
    "change": daily.diff(),
})
print(report)
weekly = daily.resample("W-SUN").sum()
print(weekly)
print("best day:", daily.idxmax().date(), daily.max())

            revenue  rolling_3d  change
date                                   
2026-01-03     41.0       41.00     NaN
2026-01-04     31.9       36.45    -9.1
2026-01-05     24.0       32.30    -7.9
2026-01-06      0.0       18.63   -24.0
2026-01-07     42.5       22.17    42.5
date
2026-01-04    72.9
2026-01-11    66.5
Freq: W-SUN, Name: total, dtype: float64
best day: 2026-01-07 42.5

Common mistakes

• Treating dates as strings: sorting and resampling don't work; parse them with parse_dates/to_datetime.
• Charting only days with data, hiding gaps; resample to a regular frequency first.
• Mixing time zones; store UTC and convert for display.
• Rolling windows without min_periods, producing NaN at the start.

Key points

• Put datetimes in the index for resampling and rolling.
• resample creates regular periods, including empty ones.
• rolling, diff and pct_change describe trends.

Further reading: Time series / date functionality · Windowing operations