Resampling and Frequency Conversion

What resampling means

Resampling converts a time series from one frequency to another:

• Downsampling — high frequency → low frequency (minute data → daily totals). This always needs an aggregation, because many points collapse into one bucket.
• Upsampling — low frequency → high frequency (weekly data → daily rows). This creates new rows, which start out empty unless you fill them.

.resample() has a similar API to .groupby(): call resample() to define the buckets, then call an aggregation like .sum() or .mean().

Downsampling: many points into one bucket

Minute data into 5-minute buckets

import pandas as pd
 
idx = pd.date_range("2024-01-01", periods=12, freq="min")
ts = pd.Series(range(12), index=idx)
 
print(ts.resample("5min").sum())

Two choices control how buckets are formed:

• closed — which edge of each interval is inclusive ("left" or "right").
• label — whether the result is labeled with the left or right edge.

closed and label options

print(ts.resample("5min", closed="right").sum())
print(ts.resample("5min", closed="right", label="right").sum())

OHLC: a finance-flavored aggregation

For price data, a common summary per bucket is open, high, low, close — computed in one call:

OHLC resampling

print(ts.resample("5min").ohlc())

Upsampling: creating new rows

Going from weekly to daily creates rows that don’t exist yet — they start as NaN unless you tell pandas how to fill them:

Upsampling with asfreq vs ffill

weekly = pd.DataFrame(
    {"a": [1, 2]},
    index=pd.date_range("2024-01-03", periods=2, freq="W-WED"),
)
 
print(weekly.resample("D").asfreq())  # new rows are NaN
print(weekly.resample("D").ffill())   # new rows carry the last known value forward

resample vs groupby

Both split data into buckets and apply a function — the difference is what defines a bucket:

• groupby("column") groups by the values in a column.
• resample("2D") groups by fixed time intervals, regardless of how many rows fall into each one.

A resample call, side by side with what groupby does

sales = pd.DataFrame(
    {"amount": [10, 20, 30, 40, 50]},
    index=pd.date_range("2024-01-01", periods=5, freq="D"),
)
 
print(sales.resample("2D").sum())

Common pitfalls

• Downsampling needs an aggregation (.sum(), .mean(), .ohlc(), …); calling .resample("5min") alone just returns a lazy resampler object, not data.
• The default for closed/label is "left" for most frequencies but "right" for a handful ("ME", "YE", "QE", "W") — always check the output on real data.
• Upsampling with .asfreq() introduces NaN; you must explicitly choose .ffill() or another fill strategy if you don’t want gaps.

Visualize it

diagram Downsampling vs upsampling mermaid

  flowchart LR
  A["High-frequency series
(e.g. 1-minute data)"] -->|"resample('5min').sum()"| B["Downsampled
fewer rows, aggregated"]
  C["Low-frequency series
(e.g. weekly data)"] -->|"resample('D').asfreq()"| D["Upsampled
more rows, mostly NaN"]
  D -->|".ffill()"| E["Upsampled + filled
no gaps"]

Downsampling aggregates many high-frequency points into fewer buckets; upsampling creates new, empty rows that need an explicit fill strategy.

sketch High-frequency points collapsing into buckets p5.js Pause Restart

Twelve minute-level points (amber) are grouped by resample into three 5-minute buckets, each collapsed to one summary bar (blue).

From the book

McKinney points out that resample’s bin edges are half-open: a data point can belong to only one interval, and the union of all intervals covers the whole time range. This is why, for the default closed="left", a value stamped exactly at 00:05:00 belongs to the next 5-minute bucket, not the one ending at 00:05:00.

🧪 Try It Yourself

Exercise 1 – Downsample Minute Data to 5-Minute Sums

pch.exerciseTag pch.exerciseTitle

Exercise 2 – Upsample and Forward-Fill

pch.exerciseTag pch.exerciseTitle

Exercise 3 – Compute OHLC in One Call

pch.exerciseTag pch.exerciseTitle

Next

Resampling collapses data into buckets. Rolling and Moving Window Functions instead slides a window across the data without collapsing rows — great for smoothing noisy series.