Time Series: resample and the dt Accessor

The monthly figures you have used since Lesson 1.3 came from somewhere: a few thousand timestamped orders, bucketed by calendar month. Time is the axis of almost every business question (“Is revenue growing?”, “Is this November normal?”, “When do people shop?”), and pandas has tools built for it. Before using them, you need one row per order with its revenue and timestamp.

import pandas as pd

items = pd.read_csv("order_items.csv")
orders = pd.read_csv("orders.csv", parse_dates=["order_date"])
items["revenue"] = items["quantity"] * items["unit_price"] * (1 - items["discount"])
order_revenue = items.groupby("order_id", as_index=False)["revenue"].sum()

orders = orders.merge(order_revenue, on="order_id", how="left", validate="one_to_one")
orders = orders[orders["status"] != "cancelled"]
print(orders[["order_id", "order_date", "revenue"]].head(3))

Every run block below starts with these lines, shortened to keep them readable. In your own notebook you would run them once.

resample: buckets of time

resample groups rows into regular time periods. It needs the dates in the index:

import pandas as pd

items = pd.read_csv("order_items.csv")
orders = pd.read_csv("orders.csv", parse_dates=["order_date"])
items["revenue"] = items["quantity"] * items["unit_price"] * (1 - items["discount"])
orders = orders.merge(items.groupby("order_id", as_index=False)["revenue"].sum(), on="order_id", validate="one_to_one")
orders = orders[orders["status"] != "cancelled"]

by_time = orders.set_index("order_date")
monthly = by_time.resample("ME")["revenue"].sum()
quarterly = by_time.resample("QE")["revenue"].sum()
print(len(monthly), "months")
print(monthly.tail(4).round(0))
print(quarterly.round(0))

The string is a frequency alias: "D" for days, "W" for weeks ending Sunday, "ME" for month ends, "QE" for quarter ends, "YE" for year ends. pandas 2.2 renamed "M", "Q" and "Y" to the E versions to make clear that the labels are period ends, and the old names now raise a deprecation warning. Periods are a different object and keep the short names: to_period("M") and to_period("Q") are correct, because a period stands for the whole month or quarter rather than its last day. The quarterly series is the one the head of growth is staring at: 80,523 dollars in Q3 2025, then 146,825 in Q4.

resample versus to_period

You can also group by a period: orders.groupby(orders["order_date"].dt.to_period("M"))["revenue"].sum(). The result looks the same here, with labels like 2025-11 instead of 2025-11-30. The difference shows when a period has no data: resample includes it with a sum of 0, while grouping skips it entirely. For charts and growth rates you want every period present, so resample is the safer default.

Comparing periods

Raw series answer “how much”; changes answer “how fast”:

import pandas as pd

items = pd.read_csv("order_items.csv")
orders = pd.read_csv("orders.csv", parse_dates=["order_date"])
items["revenue"] = items["quantity"] * items["unit_price"] * (1 - items["discount"])
orders = orders.merge(items.groupby("order_id", as_index=False)["revenue"].sum(), on="order_id", validate="one_to_one")
orders = orders[orders["status"] != "cancelled"]
by_time = orders.set_index("order_date")

quarterly = by_time.resample("QE")["revenue"].sum()
print(pd.DataFrame({
    "revenue": quarterly.round(0),
    "vs_prev_quarter": quarterly.pct_change().round(3),
    "vs_year_ago": quarterly.pct_change(4).round(3),
}))
monthly = by_time.resample("ME")["revenue"].sum()
print(monthly.rolling(3).mean().tail(3).round(0))

pct_change() divides each value by the previous one and subtracts 1. Q4 2025 is up 82% on Q3, which is the jump everyone noticed. But pct_change(4) compares each quarter with the same quarter a year earlier, and that column tells a calmer story: Q4 2025 is up 180% on Q4 2024, while Q1, Q2 and Q3 grew by 200%, 145% and 165%. Q4 grew about as fast as the rest of the year. A large part of the “jump” is a seasonal pattern that was there in 2024 too: Q4 2024 was 73% above Q3 2024.

rolling(3).mean() averages each month with the two before it, which smooths out single spikes such as November. Use it to see a trend; use the raw series to see events.

The dt accessor: calendar parts

For patterns within a period, such as days of the week or hours of the day, extract the part with .dt and group by it:

import pandas as pd

orders = pd.read_csv("orders.csv", parse_dates=["order_date"])
orders = orders[orders["status"] != "cancelled"]
d = orders["order_date"]
print(orders.groupby(d.dt.day_name()).size().sort_values())
print(orders.groupby(d.dt.hour).size().head(4))
print(d.dt.to_period("Q").value_counts().sort_index().tail(2))

Orders are spread fairly evenly across the week, from 241 on Mondays to 298 on Tuesdays and Saturdays, so there is no strong weekday effect to explain. Other useful parts: dt.year, dt.quarter, dt.month, dt.isocalendar().week, dt.date and dt.normalize() (midnight of the same day, which you used in Lesson 3.4).

Daily resampling is where assumptions get tested. by_time.resample("D").size() gives one count per calendar day, 729 of them from the first order to the last, and two facts fall out at once: 119 days had no orders at all (in a series built with groupby, those days would simply be missing), and the busiest single day was 10 November 2025 with 16 orders, almost three weeks before Black Friday (28 November). The other top days, 6, 8 and 26 November, are spread across the month too. Whatever drove November, it was not one big promotional day; keep that in mind for the case study.

In the exercise you build the monthly series and the year-over-year view yourself. Next you rank and bin customers, turning a continuous measure like total spend into segments.