Zansa › Answers › Forecasting and data science
How is time-series cross-validation different from normal cross-validation?
Time-series cross-validation always trains on the past and tests on the future, moving the cutoff date forward in steps. Normal cross-validation shuffles rows randomly, which lets a model see future data while predicting the past. For forecasting that leakage gives overoptimistic results, so ordering must be respected.
In a rolling-origin scheme you might train on months 1 to 24 and test on month 25, then train on 1 to 25 and test on 26, and so on. Alternatively you can use a fixed-size window that slides forward, which helps if old data is no longer representative.
Average the errors across all folds, and look at how they vary, since a method that is good on average but occasionally terrible may be risky. Allow a gap between training and test if features use lagged information.
Want this applied to your business? Zansa builds dashboards and forecasts for small and mid-sized companies worldwide. Start with a free intro call.
Book a call