Chapter 17 Dates and Basic Time-Series Operations

Dates require explicit types and ordering. String dates may sort incorrectly, and time-based calculations are unreliable when rows are unsorted or time zones are mixed.

17.1 Create and parse dates

import pandas as pd

daily_index = pd.date_range(start="2026-01-01", periods=4, freq="D")

assert daily_index[0] == pd.Timestamp("2026-01-01")
assert daily_index.freqstr == "D"

Use an explicit format for known input formats:

raw_dates = pd.Series(["07/01/2026", "07/15/2026"])
parsed_dates = pd.to_datetime(raw_dates, format="%m/%d/%Y")

assert parsed_dates.dt.month.tolist() == [7, 7]
assert parsed_dates.dt.day.tolist() == [1, 15]

17.2 Sort before calculating changes

diff() computes the change from the previous row in the current order. Sort by entity and time before using it on panel data.

revenue = pd.DataFrame(
    {
        "date": pd.to_datetime(["2026-01-01", "2026-01-02", "2026-01-03"]),
        "revenue": [100.0, 120.0, 90.0],
    }
).sort_values("date")

revenue["daily_change"] = revenue["revenue"].diff()
revenue["growth_rate"] = revenue["revenue"].pct_change()

assert pd.isna(revenue.loc[0, "daily_change"])
assert revenue.loc[1, "daily_change"] == 20.0
assert round(revenue.loc[1, "growth_rate"], 2) == 0.20

The first difference is missing because no prior observation exists. Decide whether to retain that missing value or define a domain-specific baseline; do not automatically replace it with zero.

17.3 Grouped differences

For multiple entities, calculate changes within each entity rather than across adjacent rows belonging to different groups.

panel = pd.DataFrame(
    {
        "customer_id": [1, 1, 2, 2],
        "date": pd.to_datetime(["2026-01-01", "2026-02-01"] * 2),
        "spend": [10, 15, 20, 18],
    }
).sort_values(["customer_id", "date"])

panel["spend_change"] = panel.groupby("customer_id")["spend"].diff()

assert panel["spend_change"].tolist()[1] == 5
assert panel["spend_change"].tolist()[3] == -2

17.4 Resampling requires a time index

daily = pd.Series(
    [10, 20, 30, 40],
    index=pd.date_range("2026-01-01", periods=4, freq="D"),
)
two_day_totals = daily.resample("2D").sum()

assert two_day_totals.tolist() == [30, 70]

Aggregation frequency changes the unit of analysis. State whether a result is a daily sum, monthly average, end-of-period value, or another defined measure.

17.5 Interview checklist

Are values true datetimes or only strings?
Are entity and time ordering explicit?
Should differences be calculated within groups?
What does the first missing difference mean?
Does resampling use the correct aggregation and frequency?
Could future information enter a feature through a centered or forward-looking window?

17.6 Exercises

Calculate month-over-month revenue change for two products.
Explain why applying diff() before sorting can produce plausible but wrong results.
Create a seven-day rolling mean using only current and past observations.