This is a little bit complicated topic. When you have cross-sectional data time series, training validation holdout split is a reasonable thing one could do. First of all one must set aside or spend little time for validating data. Training the candidate models is mandatory. Using the model’s fir quality that determines validation data is one of the best options. Then collecting data between validation & current data is also mandatory. Ultimately train model before the final holdout period. Holdout data can be used to compute the model.
Model quality is more variable & reliable as compared to cross-sectional data model. There is a lot of auto correction within the validation data set. A good solution suggested can be backtesting with many validation periods. One can make model on year 1 & compute the performance in the year 2, using the 2 model or both mixed & compute the performance in year 3. Almost 4 validation periods to be utilised as well as compute the performance. Also there are various amount of validation time series models but the main ones are: setting up validation data model for a time period & holdout period. Building data model aa well as computing test performance in consecutive years. A multi-fold variation of the first option.