Fenêtres glissantes et décalages

Raw time series can be jumpy. A rolling window averages nearby points so the underlying signal becomes easy to see. 🌊

A window is a fixed number of recent points. It slides forward one step at a time, recomputing a summary at each position.

The most common move is a rolling mean: average the last few values to smooth the line. pandas makes it a one-liner.

A larger window smooths more but reacts slowly; a smaller one stays responsive but noisy. You tune it to the question.

A window of 7 needs 7 points before it can compute. So the first rows come back as NaN until enough history exists.

Rolling works with more than averages. A rolling std reveals how volatility changes over time, which trends alone hide.

A lag is simply the value from an earlier time step. Yesterday's sales as a column is a lag-1 feature.

pandas builds lags with shift. Calling shift(1) pushes every value down one row so each row sees its past.

Most series depend on their own recent past. Feeding a model good lag features lets it learn that yesterday predicts today.

Subtracting the previous value, value minus its lag, gives the change between steps. It is the seed of differencing.

Strong feature sets often mix both: a few lags plus a rolling mean. Together they capture recent level and direction.

Which pandas method shifts values to their previous time step to build a lag feature?

You used rolling windows to smooth and lags via shift to bring the past forward, the two staples of time series features.