Stop 11 · Time Series AnalysisTopics · 6/7

Time Series Cross-Validation

In short

Shuffling rows into random folds lets a forecasting model peek at the future. Time series cross-validation keeps time in order, so your backtest matches how forecasts are really used.

Standard k-fold cross-validation doesn't work for time series (it would leak future data into training) — instead, you use expanding or rolling windows that always train on the past and validate on the future.

The mascot works through a row of old exam papers stacked oldest to newest, with a sealed envelope at the end.

Think of it like this

Practising with past papers in order

To predict what this year's board paper looks like, you study the 2019 to 2021 papers and test yourself on 2022. Then you add 2022 and test on 2023. You never use 2024's paper to 'predict' 2023. Time only runs forwards.

Watch it happen

Expanding-window folds over 10 months

Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

1/5Ten months of data in time order. Each fold trains on the past and tests on the next stretch.

The big ideas

What you need to know

Train past, test future

Each fold respects time order. No future rows ever reach training.

Expanding or sliding

Expanding windows keep all history. Sliding windows keep a fixed length, useful when old data is stale.

Match the horizon

If you forecast 2 weeks ahead in production, make each test window 2 weeks long.

Watch for leaks

Lag features and rolling averages must be built only from data before each prediction date.

See it in code

Walk through the code

python

1import numpy as np2from sklearn.model_selection import TimeSeriesSplit3 4X = np.arange(10).reshape(-1, 1)    # 10 months in order5 6tscv = TimeSeriesSplit(n_splits=3, test_size=2)7for fold, (train_idx, test_idx) in enumerate(tscv.split(X)):8    print(fold, "train:", train_idx, "test:", test_idx)

1/3Ten rows standing in for ten months, already sorted by date.

Prints the same three folds as the scene: train 0 to 3 then test 4 and 5, and so on.

Save yourself the pain

Common mistakes

Oops

Using KFold with shuffle=True on time-ordered data.

Do this instead

Use TimeSeriesSplit, or manual date-based cut-offs, so tests are always in the future.

Oops

Computing a rolling mean over the whole series before splitting.

Do this instead

Build lag and rolling features inside each fold, or with shift() so they only use the past.

Your mission

Random vs time-ordered folds

~30 min

Quick check

Why is shuffled k-fold cross-validation risky for forecasting?

Go deeper

1Implement a rolling-window cross-validation split on a sample time series.
2Learn why randomly shuffling time series data before splitting is a critical mistake.
3Compare model performance across multiple rolling windows, not just a single split.