Definition
- A blocked split (also called single holdout for time series) is the simplest form of time series validation.
- The dataset is split into one contiguous block for training and one later contiguous block for testing.
- Unlike random splits, this method respects temporal order (train on past → test on future).
How It Works
- Arrange dataset in chronological order.
- Select an early block as the training set.
- Select the later block as the validation/test set.
- Train model on train block → evaluate on test block.
Example
Dataset: 2018–2022
- Train = 2018–2020
- Test = 2021–2022
Single cut → one train/test split.
Why It’s Used
- Simple and fast.
- Often the first baseline evaluation for time series models.
- Useful when retraining frequently isn’t necessary or data is stable.
Limitations
- Only one split → evaluation may be sensitive to choice of cutoff point.
- Doesn’t give robust generalization estimates (compared to CV methods like expanding/sliding windows).
- If test period is unusual (e.g., pandemic spike), results may misrepresent model performance.
Comparison with Other Methods
| Method | Splits | Train Set | Test Set | Robustness |
|---|---|---|---|---|
| Blocked (Single Holdout) | 1 | Single block (past) | Single block (future) | Low |
| Expanding Window | Many | Grows over time | Moves forward | Medium |
| Sliding Window | Many | Fixed size, rolling | Moves forward | High |
Summary
Blocked Split (Single Holdout) = splitting time series data once into past (train) and future (test) blocks.
- Simple, fast, respects time order.
- But less robust because it uses only one test split.