Cross-validation for time series is a method used to assess the predictive performance of models while accounting for temporal dependencies in the data. It involves techniques like rolling forecasting origin and time series split to ensure that the training and test sets respect the chronological order of observations.