Rolling and Moving Window Functions

Why moving windows matter

Real time series are noisy — daily sales bounce around, stock prices jitter minute to minute. A moving window function recomputes a statistic (mean, std, …) over a small sliding slice of the data, smoothing out that noise so the underlying trend is easier to see.

Pandas gives you three flavors:

• .rolling(window) — a fixed-size window that slides forward.
• .expanding() — a window that starts small and grows to include everything seen so far.
• .ewm(span=...) — an exponentially weighted window that favors recent observations.

Rolling mean and std

Rolling window basics

import pandas as pd
 
s = pd.Series(
    [10, 12, 13, 12, 15, 20, 18, 22, 25, 30],
    index=pd.date_range("2024-01-01", periods=10, freq="D"),
)
 
print(s.rolling(3).mean())  # 3-day moving average

rolling(3) behaves like groupby, but instead of grouping by a key, it creates a sliding 3-row window. Notice the first two values are NaN — there aren’t 3 rows yet to average.

min_periods: dealing with a short warm-up

By default, a rolling window needs every slot filled before it produces a value. min_periods relaxes that, useful right at the start of a series:

min_periods relaxes the warm-up requirement

print(s.rolling(3, min_periods=1).mean())  # no leading NaN this time

Expanding windows

An expanding window is like a rolling window whose size keeps growing — it always starts at the beginning of the series and includes everything up to the current row:

Expanding mean

print(s.expanding().mean())  # cumulative average up to each point

Exponentially weighted windows (ewm)

Instead of weighting every value in the window equally, ewm gives more weight to recent observations — it “reacts” faster to a change than a simple moving average:

Exponentially weighted mean

print(s.ewm(span=3).mean())

Rolling correlation (two series at once)

Some window functions — like correlation — need two series. Calling .rolling(...) on one series and then .corr(other) computes a rolling correlation between them:

Rolling correlation between two return series

bench = pd.Series(
    [1, 1.2, 1.1, 1.3, 1.5, 1.6, 1.4, 1.7, 1.9, 2.0],
    index=s.index,
)
 
rolling_corr = s.pct_change().rolling(4, min_periods=2).corr(bench.pct_change())
print(rolling_corr)

Common pitfalls

• Without min_periods, the first window - 1 rows of a rolling result are always NaN.
• Rolling window functions automatically skip missing data, but they still require min_periods valid values to produce output.
• ewm doesn’t have a fixed “window size” the way rolling does — span just controls how quickly old observations lose influence.

Visualize it

diagram Three ways to slide a window mermaid

  flowchart LR
  A["Raw noisy series"] --> B["rolling(3).mean()
fixed-size sliding window"]
  A --> C["expanding().mean()
grows from the start"]
  A --> D["ewm(span=3).mean()
recent points weighted more"]
  B --> E["Smoothed line"]
  C --> E
  D --> E

rolling keeps a fixed-size window; expanding grows from the start; ewm weights recent points more heavily instead of using a hard cutoff.

sketch A rolling window smoothing a noisy series p5.js Pause Restart

The amber line is the raw noisy series; the blue line is its 3-point rolling mean, visibly smoother.

From the book

McKinney frames rolling(250) on daily stock prices as computing a “250-day moving window average” — the same mental model as groupby, except the groups are sliding windows instead of category labels. He also notes that moving window functions automatically exclude missing data, just like other pandas statistical methods.

🧪 Try It Yourself

Exercise 1 – Compute a Rolling Mean

pch.exerciseTag pch.exerciseTitle

Exercise 2 – Avoid Leading NaN With min_periods

pch.exerciseTag pch.exerciseTitle

Exercise 3 – Compute an Expanding Mean

pch.exerciseTag pch.exerciseTitle

Next

That completes the Pandas phase. Head into Phase 4 — Data Preprocessing and Cleaning, starting with Understanding Data Quality, to apply these skills to real, messier datasets.