What are Time-Based Splits?

• Time-based splits respect the chronological order of data.
• Unlike random splits, you never train on future data to predict the past.
• This is crucial for time series forecasting and any task where data has a temporal component (financial prediction, demand forecasting, sensor logs, etc.).

Key principle:

$\text{Training data time range } < \text{Validation data time range}$

Why not use K-Fold?

• Standard CV (random or stratified) shuffles data → mixes past & future.
• This leads to data leakage: the model “sees” the future during training.
• Example: If you’re forecasting sales for July, you can’t train on August data.

Common Time-Based CV Strategies

1. Expanding Window (Growing Training Set)

• Training set grows over time, test set is the next fixed chunk.
• Example:
  • Fold 1: Train [Jan–Mar], Test [Apr]
  • Fold 2: Train [Jan–Apr], Test [May]
  • Fold 3: Train [Jan–May], Test [Jun]

2. Sliding Window (Rolling Window)

• Training set is a fixed-length moving window.
• Example (window = 3 months):
  • Fold 1: Train [Jan–Mar], Test [Apr]
  • Fold 2: Train [Feb–Apr], Test [May]
  • Fold 3: Train [Mar–May], Test [Jun]

3. Blocked Splits (Single Holdout)

• Simple chronological split: train on first N%, test on last (1–N)%.
• Often used for initial baseline.

Example in Python (sklearn TimeSeriesSplit)

import numpy as np
from sklearn.model_selection import TimeSeriesSplit

X = np.arange(10).reshape(-1, 1)
y = np.arange(10)

tscv = TimeSeriesSplit(n_splits=3)

for train_idx, test_idx in tscv.split(X):
    print("Train:", train_idx, "Test:", test_idx)

import numpy as np
from sklearn.model_selection import TimeSeriesSplit

X = np.arange(10).reshape(-1, 1)
y = np.arange(10)

tscv = TimeSeriesSplit(n_splits=3)

for train_idx, test_idx in tscv.split(X):
    print("Train:", train_idx, "Test:", test_idx)

Output (example):

Train: [0 1 2 3]       Test: [4 5]
Train: [0 1 2 3 4 5]   Test: [6 7]
Train: [0 1 2 3 4 5 6 7] Test: [8 9]

Train: [0 1 2 3]       Test: [4 5]
Train: [0 1 2 3 4 5]   Test: [6 7]
Train: [0 1 2 3 4 5 6 7] Test: [8 9]

Here:

• Training expands with each fold.
• Tests are strictly after the training window.

When to Use

Use time-based splits when:

• Predicting future from past (time series forecasting, stock predictions).
• Data has seasonality or temporal autocorrelation.
• Avoiding data leakage is critical.

Not necessary when:

• Data points are IID (independent and identically distributed).
• Labels don’t depend on time order.

Summary

• Time-based splits ensure temporal order is respected.
• Prevents training on future data.
• Implemented as expanding window, sliding window, or blocked holdout.