Parts of a Date: .dt
Once a column holds dates, .dt gives its parts for every row at once, such as year, month, day_name(), hour and date, so you can count by weekday or hour, and subtracting two dates gives a Timedelta, the gap between them, whose .dt.days counts whole days.
Apprentice · Python · Updated
After this lesson you can
- Pull parts out of a datetime column with .dt, such as year, month, day_name(), hour and date
- Work out the gap between two dates as a Timedelta, and count rows by weekday in calendar order
Key takeaways
- .dt, read "dot d t", gives a date column its date tools, as .str gives a text column its string tools. It works only on a datetime64 column.
- .dt.year, .dt.month, .dt.hour and .dt.date take one part of every date, and .dt.day_name() gives the day of the week as a word.
- value_counts() on a part counts the rows for each day or hour, and reindex() puts the days in calendar order.
- Subtracting one date from another gives a Timedelta, the length of time between them, and .dt.days gives it in whole days.
Which day do most orders arrive? What hour is the help desk busiest? How many days before a deadline do people sign up? Questions like these come up in every job that keeps dates, and they're all about one part of a date, or the gap between two. This lesson takes a date column apart.
The example is the Hillcrest 5K Fun Run's registrations, registrations.csv: 450 runners and the moment each sent the sign-up form, registered_at, such as 2026-08-03 07:40. The race is on Saturday, October 3, 2026, and the race director is planning next year's advertising: on which days, and at what time, do people sign up? The runners and the race are our own made-up data.
.dt: the date tools for a whole column
Load the file with registered_at read as dates, as Parsing Dates with to_datetime shows:
import pandas as pd
reg = pd.read_csv("registrations.csv", parse_dates=["registered_at"])
signed = reg["registered_at"]
signed.head(3)
0 2026-08-03 07:40:00
1 2026-08-04 06:27:00
2 2026-08-04 06:49:00
Name: registered_at, dtype: datetime64[us]
Each value is a Timestamp, pandas' word for one date and time. A date column has a set of tools under .dt (read "dot d t"), as a text column has .str, and each works on every row at once. A few of them take one part of each date:
pd.DataFrame({
"registered_at": signed,
"year": signed.dt.year,
"month": signed.dt.month,
"day": signed.dt.day_name(),
"hour": signed.dt.hour,
}).head(3)
registered_at year month day hour
0 2026-08-03 07:40:00 2026 8 Monday 7
1 2026-08-04 06:27:00 2026 8 Tuesday 6
2 2026-08-04 06:49:00 2026 8 Tuesday 6
year, month and hour are numbers, with no brackets, because they're stored values. day_name() has brackets because it works out a word for each day. .dt works only on a datetime64 column: on text such as "2026-08-03 07:40", pandas stops with an error that says so.
Counting by weekday
With the weekday as a word, value_counts() counts the registrations for each day:
signed.dt.day_name().value_counts()
registered_at
Tuesday 141
Thursday 67
Saturday 64
Wednesday 60
Monday 47
Friday 45
Sunday 26
Name: count, dtype: int64
Tuesday stands out, with more than twice any other day. value_counts sorts from most to fewest, which is right for "which day is busiest", but a week reads better in calendar order. reindex() puts the rows in the order of a list you give it, and fill_value=0 writes 0 for a day that isn't there at all, instead of NaN:
week = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"]
signed.dt.day_name().value_counts().reindex(week, fill_value=0)
registered_at
Monday 47
Tuesday 141
Wednesday 60
Thursday 67
Friday 45
Saturday 64
Sunday 26
Name: count, dtype: int64
Hours and days
Why Tuesday? Take only Tuesday's registrations and count them by hour:
tuesdays = signed[signed.dt.day_name() == "Tuesday"]
tuesdays.dt.hour.value_counts().head(3)
registered_at
20 53
19 25
21 16
Name: count, dtype: int64
Most Tuesday sign-ups come between 7 and 10 in the evening. Hillcrest's running club meets on Tuesday evenings, so runners register when they get home from the club run. That's next year's advertising plan: a poster at the club, and a reminder on Tuesday nights.
.dt.date drops the time and keeps the calendar day, which counts registrations per day:
signed.dt.date.value_counts().head(3)
registered_at
2026-08-04 22
2026-09-15 20
2026-09-01 18
Name: count, dtype: int64
The three busiest days are all Tuesdays, the first one the day after the form opened.
The gap between two dates: Timedelta
How long before the race did people sign up? Subtract each registration from race day. Subtracting one date from another gives a Timedelta, a length of time, such as "60 days and 16 hours":
race_day = pd.Timestamp("2026-10-03")
before = race_day - signed
before.head(3)
0 60 days 16:20:00
1 59 days 17:33:00
2 59 days 17:11:00
Name: registered_at, dtype: timedelta64[us]
A Timedelta column has .dt tools too. .dt.days gives the whole days, which makes a question such as "who signed up in the last week?" a comparison:
print("longest wait:", before.max())
print("signed up in the last week:", (before.dt.days < 7).sum())
longest wait: 60 days 16:20:00
signed up in the last week: 50
Fifty runners signed up in the final week, so next year's form stays open until the night before. A Timedelta also measures how long a run took, and Durations: Finish Times with Timedelta uses it for that.
Check yourself
2 questions to try on the lesson page.
So what
.dt turns a date column into the parts your question is about, a weekday, an hour, a day, and subtracting dates gives a Timedelta you can measure in days, so "when?" and "how long before?" become counts you can act on.