Asked about Data where order carries meaning
How do you build a rolling average feature safely?
An answer that works
Shift before rolling, so the window covers only past values. A rolling mean that includes the current row leaks the target and produces a model that cannot work in production.
Reading an answer is not being able to give one. The section that teaches this is Data where order carries meaning, 5 hours of Time series.
Learn this properlyAlso asked about Time series
Why can you not use ordinary cross-validation on a time series?What is the first thing you do on a forecasting problem?Would you use a neural network or a classical model for forecasting?Why evaluate at multiple forecast horizons?How would you evaluate a demand forecasting model?You need forecasts for 50,000 products, many with little history. How do you approach it?