Time-Series Splitting

Time-ordered data breaks the i.i.d. assumption underlying standard cross-validation. The fundamental rule is simple: always train on the past, evaluate on the future. Any split that allows information from a future time point to influence training is a data leak that will inflate reported performance.

DataSplits provides four strategies for time-ordered data. They differ in how strictly they enforce temporal order and how they handle the label-overlap leakage that arises in financial and event-based data.

The atomicity rule

All time-series strategies share one convention: observations sharing the same timestamp are never split between train and test in the same fold. Block and chunk boundaries always fall between distinct time values. This prevents partial-group leakage when multiple observations carry the same timestamp (e.g. daily or weekly data with multiple instruments).

TimeSplit — single train/test cutoff

TimeSplit is the simplest strategy: sort observations chronologically and put the first train observations into the training set, the rest into test.

using DataSplits

# Oldest 80% of observations go to train.
res = partition(X, TimeSplit(:asc); time = timestamps, train = 0.8, test = 0.2)
X_train, X_test = splitdata(res, X)

# Convenience aliases.
res = partition(X, TimeSplitOldest(); time = timestamps, train = 0.8, test = 0.2)
res = partition(X, TimeSplitNewest(); time = timestamps, train = 0.2, test = 0.8)

# When the timestamps vector is the data.
res = partition(timestamps, TimeSplitOldest(); train = 0.8, test = 0.2)

TimeSplit(:asc) places the oldest observations in train; TimeSplit(:desc) places the newest in train (useful for testing on historical data).

The actual train fraction may slightly overshoot train because complete timestamp groups are never broken.

TimeSeriesSplit — expanding or rolling window CV

TimeSeriesSplit produces k folds in chronological order. Fold i trains on the observations before chunk i+1 and tests on chunk i+1. By default the training window grows across folds (expanding window); set max_train_size for a rolling window of fixed length.

using DataSplits

# Expanding window: each successive fold trains on more data.
cvs = partition(X, TimeSeriesSplit(5); time = timestamps)

for (X_tr, X_te) in splitview(cvs, X)
    fit!(model, X_tr)
    evaluate(model, X_te)
end

# Rolling window: train on at most the last 200 observations.
cvs = partition(X, TimeSeriesSplit(5; max_train_size = 200); time = timestamps)

# Add a gap between the end of train and the start of test to avoid leakage
# from features that look ahead (e.g. moving averages).
cvs = partition(X, TimeSeriesSplit(5; gap = 5); time = timestamps)

Use TimeSeriesSplit when you have enough data for expanding or rolling evaluation and you want a standard walk-forward backtesting setup.

BlockedCV — test block surrounded by train

BlockedCV (Bergmeir & Benítez 2012, Roberts et al. 2017) divides the data into k contiguous blocks. Each block takes a turn as the test set while all other blocks — both before and after — form the training set. A symmetric gap is removed on both sides of the test block to buffer against autocorrelation.

cvs = partition(X, BlockedCV(5); time = timestamps)

# Remove 2 observations on each side of the test block.
cvs = partition(X, BlockedCV(5; gap = 2); time = timestamps)

This differs from TimeSeriesSplit in a key way: later blocks can appear in the training set of earlier test folds. This is valid when the data is stationary and what you want is average-in-time performance rather than forward-only prediction.

PurgedKFold — purging and embargo

PurgedKFold is the standard cross-validation strategy for financial machine learning (López de Prado 2018). It extends blocked CV with an asymmetric gap:

• Purge (purge observations): removed from train before the test block, to prevent leakage from samples whose labels overlap the test period (e.g. returns computed from forward-looking windows that reach into the test window).
• Embargo (embargo observations): removed from train after the test block, to prevent leakage from serial correlation between test-period features and the immediately subsequent train samples.

# 5-fold purged CV with a 2-observation purge and 1-observation embargo.
cvs = partition(X, PurgedKFold(5; purge = 2, embargo = 1); time = timestamps)

for (X_tr, X_te) in splitview(cvs, X)
    fit!(model, X_tr); evaluate(model, X_te)
end

# Minimal version — no purge or embargo (equivalent to BlockedCV).
cvs = partition(X, PurgedKFold(5); time = timestamps)

Choosing between the four strategies

References

• Bergmeir, C.; Benítez, J. M. On the use of cross-validation for time series predictor evaluation. Inf. Sci. 2012, 191, 192–213.
• Roberts, D. R. et al. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 2017, 40(8), 913–929.
• López de Prado, M. Advances in Financial Machine Learning. Wiley, 2018, §7.4.