29 / 97
Embargo Period
Working definition
A span of observations discarded immediately after a test window, so that serial correlation cannot carry information from the tested period into the data used to train.
Purging removes training observations whose information intervals overlap a test window. An embargo removes the ones that follow it.
The reason is serial correlation. Prices, volatilities, spreads and most engineered features are correlated with their own recent past, so an observation drawn from the hours or days immediately after a test fold is not independent of that fold — it carries a diluted copy of the same information. Train on it and the model has been handed a hint about the period it will be graded on, even though no interval formally overlaps. The leak is smaller than an unpurged overlap and it points the same way: validation scores that flatter.
The embargo is stated as a fraction of the sample or as a fixed span, and the defensible way to choose it is from the data’s own structure — the horizon over which feature autocorrelation decays to negligible, plus the label horizon. Chosen after seeing which value produces the nicest results, it becomes one more researcher degree of freedom, which is why the span belongs in a pre-registration written before the split is run.
Note the asymmetry. Purging applies on both sides of the test window; an embargo applies only after it. Training data preceding the test period needs no embargo, because the leakage that matters flows forwards in time, out of the tested period and into the training set.
Why it matters
Embargoing is cheap — it discards a narrow band of rows — and it closes the last routine channel by which a time-series split can grade a model on data it has partly seen. Together with purging it lets a cross-validated result on market data carry the meaning a cross-validated result carries anywhere else: that the score was earned on observations the model had no access to.
Commonly confused with
Neighbouring concepts that get used interchangeably, and the distinction that actually separates them.
- Purging
Purging removes training observations whose information intervals overlap a test window. An embargo removes the ones that follow it. Note the asymmetry — purging applies on both sides, an embargo only after, because the leakage that matters flows forwards out of the tested period.
- A gap between train and test
A fixed gap chosen for tidiness and an embargo chosen from the data's structure look identical in code and are not the same claim. The defensible span comes from the horizon over which feature autocorrelation decays to negligible, plus the label horizon.
- A holdout period
A holdout is data reserved for evaluation. An embargo is data discarded entirely — used by neither side of the split. It buys nothing except the removal of a leak, which is exactly why it is cheap and easy to skip.
- Effective sample size
Both respond to the same serial correlation. Effective sample size measures how much independent evidence the data holds; an embargo removes the specific rows through which that dependence would leak across a split. One is a measurement, the other a remedy.
How to measure it in your own data
A definition you cannot test is a definition you have to take on trust. This is the shortest honest route from the concept to a number you computed yourself.
- Records you need
The autocorrelation structure of the features and the forward horizon of the label. Both are properties of the data rather than choices, which is what makes a principled span available.
- What you compute
Set the span from the horizon over which feature autocorrelation decays to negligible, plus the label horizon, and apply it to the observations immediately following each test window. State it as a fraction of the sample or a fixed span, and record which.
- What the answer tells you
Choose the span before running the split. Chosen after seeing which value produces the nicest results it becomes one more researcher degree of freedom, which is precisely the failure the whole apparatus exists to prevent — so the span belongs in a pre-registration rather than in the write-up.
Questions and answers
Why does the embargo only apply after the test window?
Because the leakage that matters flows forwards in time, out of the tested period and into the training set. Training data that precedes the test window cannot have been contaminated by outcomes that had not happened yet, so it needs no embargo. The asymmetry is a consequence of causality rather than a convention.
How long should the embargo be?
Long enough that the dependence has decayed: the horizon over which feature autocorrelation becomes negligible, plus the label horizon. That gives a span you can defend from the data. Any span chosen because it produced a better score is a researcher degree of freedom wearing a technical name.
Is an embargo worth the data it costs?
Almost always, because it costs very little. It discards a narrow band of rows immediately after each test fold, and in exchange it closes the last routine channel by which a time-series split can grade a model on data it has partly seen.
What does embargoing plus purging actually buy me?
Embargoing and purging buy one thing: that a cross-validated result on market data carries the meaning a cross-validated result carries anywhere else — that the score was earned on observations the model had no access to. Without both, the number resembles a validation score without being one.
Related terms
Derived from the links this entry makes and the entries that link back to it.
In the research
Embargo Period comes up in one research note on this site.