TimeSeriesSplit

TimeSeriesSplit is a cross-validation strategy used for time-ordered data. It is designed to preserve the chronological order of observations and to prevent leakage of future information into training sets, which can bias performance estimates in forecasting tasks. It is available as TimeSeriesSplit in scikit-learn's model_selection module.

How it works: The data is partitioned into n_splits+1 consecutive blocks of roughly equal size. Each split selects the training set as all observations up to the start of the next block, and the test set as the observations in that next block. As a result, the training set grows with each successive split, while the test set remains a fixed, contiguous block. No shuffling is performed.

Parameters and usage: The primary parameter is n_splits, the number of train/test splits to generate. Optional parameter max_train_size limits how large the training portion may grow in any split. The split method yields train_index and test_index for each split, suitable for indexing into X and y. A typical usage pattern is to create a TimeSeriesSplit instance and loop over the resulting indices to evaluate a model across rolling-origin forecasts.

Notes: TimeSeriesSplit is particularly suitable for evaluating models where future data must not contaminate the training set, such as stock prices or weather forecasts. It may leave trailing observations unused depending on data length and split configuration. It complements other cross-validation styles that do not respect temporal order.