A model is trained on data from 2020–2022 and tested on data from 2023. What type of split is this and why is it appropriate for time series?