Filling Gaps in Time Series

Time series data often has gaps: a sensor fails, a survey is skipped, or a fiscal year's report is missing. Three methods—forward fill, backward fill, and interpolation—each make different assumptions about how the missing value relates to its neighbours. The choice matters for prediction and causal inference.

The idea

Forward fill (ffill) copies the last known value forward. It assumes the variable stays constant until new data arrives. Backward fill (bfill) copies the next value backward. It assumes the future value existed earlier, which leaks future information into the past and breaks temporal causality. Interpolation—linear, spline, or polynomial—fits a curve between neighbours. It assumes smooth transitions.

For prediction tasks, forward fill is safe: the model sees only past values. Backward fill is not safe: it uses future values the model would not have at prediction time. Interpolation is safe if it uses only past neighbours, dangerous if it uses future ones.

The maths

Try it

Method:

Original

∅

ffill

Method: ffill

Interactive demonstration of time series concepts

Where I used it

Easy to get wrong

Sources

MAST90083 Time Series (2024). Missing data strategies for forecast models.