Dates and Ordered Data

Rolling Mean (Window = 3)

Compute a 3-element sliding-window average over an ordered numeric sequence. By hand, the first two positions cannot form a full window so they get None; positions 2 onward average the current element and the two before it. With pandas, Series.rolling(3).mean() handles the window and fills the incomplete prefix with NaN.

By hand

With pandas

df['x'].rolling(3).mean() computes the mean of each 3-row window in one call. The first two rows have no complete window, so they are NaN and the column is float64. The snapshot shows the raw output; result converts NaN → None and rounds the finite values for comparison with the naive half.

naive.py 
readings = [6, 9, 12, 9, 15, 9]
result = []
for i in range(len(readings)):
    if i < 2:
        result.append(None)
    else:
        w = readings[i - 2: i + 1]
        result.append(round(sum(w) / 3, 2))
print('RESULT:', result)
library.py 
import math
import pandas as pd
from dalib.display import set_display
set_display()

readings = [6, 9, 12, 9, 15, 9]
df = pd.DataFrame({'x': readings})
r = df['x'].rolling(3).mean()
result = [round(v, 2) if not math.isnan(v) else None for v in r.tolist()]
print('index:', r.index.tolist())
print('dtype:', r.dtype)
print('values raw:', r.tolist())
print('RESULT:', result)

index: [0, 1, 2, 3, 4, 5]
dtype: float64
values raw: [nan, nan, 9.0, 10.0, 12.0, 11.0]
RESULT: [None, None, 9.0, 10.0, 12.0, 11.0]

Implementation notes

  • The min_periods parameter controls how many non-NaN values are required to compute a result. rolling(3, min_periods=1) fills the prefix instead of returning NaN.
  • rolling also supports sum(), min(), max(), std(), and apply(fn) for custom window functions.
  • The float64 upcast and NaN prefix mirror diff() — use math.isnan or pd.isna when normalising for comparison.
  • Cross-reference: diff-previous-row (this chapter) for a single-step lag instead of an aggregated window.
  • Cross-reference: running-total (python-data-basics) for the cumulative-sum approach — accumulating all prior values rather than a fixed-size window.