A data engineer needs to split time-series data for training a forecasting model. The data is sorted by timestamp. The engineer wants to avoid leakage where future data influences training. Which data splitting approach should they use?