How do rolling and expanding windows work in pandas, and when do you use each?
rolling() computes statistics over a fixed-size sliding window, discarding data outside the window; expanding() grows the window from the first row to the current row, equivalent to an ever-increasing cumulative calculation. Both return objects you chain .mean(), .sum(), .std(), or a custom .apply() onto.
How to think about it
How to think through window functions
The key distinction is what you’re forgetting versus what you’re remembering. A rolling window is like a sliding spotlight — it only looks at the most recent N rows and actively forgets older data. An expanding window is like a running total — it never forgets anything, growing wider with every new row.
When does that matter? Rolling is right for short-term trend detection: is this week’s sales higher than the 4-week average? Expanding is right for running metrics: what is the all-time average up to today?
Try rolling and expanding side by side
Time-based windows on a DatetimeIndex
When your index is a DatetimeIndex, you can pass a time offset string instead of a row count. This handles irregular time series correctly — a “3D” window always covers 3 calendar days regardless of how many rows fall in that window.
ts = prices.copy()
ts.index = pd.date_range("2024-01-01", periods=10, freq="D")
# 3-calendar-day rolling window
ts.rolling("3D").mean()GroupBy + rolling for per-group moving averages
This combination is one of the most common feature engineering patterns in time series ML:
df["rolling_avg"] = (
df.groupby("ticker")["close"]
.transform(lambda s: s.rolling(window=5).mean())
)The transform ensures the result is index-aligned with the original DataFrame, so you can assign it directly as a new column without a merge.
Custom functions on a window
# 3-period price momentum (last minus first in window)
prices.rolling(3).apply(lambda x: x[-1] - x[0], raw=True)raw=True passes a NumPy array instead of a Series to the lambda, skipping pandas overhead — significantly faster for large windows.