Validation method

Combinatorial purged cross-validation

CPCV · not a gate on Alphavo

Combinatorial purged cross-validation rotates which blocks of a time series are held out, and purges labels that overlap the test.

What CPCV is

A single cut of a backtest into "train" and "test" can flatter a rule that happened to fit that cut. Combinatorial purged cross-validation, the design associated with López de Prado's work on backtest overfitting, splits the series into blocks and rotates which blocks are the test. Many combinations are scored, not one favorite split. Purging removes training observations whose labels reach into a test block, so the fit is less able to see the test. The question is whether the result depended on one way of cutting the sample.

How Alphavo uses the name

Alphavo does not use CPCV as a gate on the path that shows a strategy. In-Sample Testing chooses parameters and scores them on the research sample. Out-of-Sample Testing reads signal edge on an expanding window. Robust Walk-Forward rolls the exit check. T6 is in-sample calendar-quarter stability, on its own book. Stress raises slippage and funding, then looks at cash. The sandbox replays the account on historical prices. CPCV is not that path.

What CPCV does not set

A shown strategy is 0 stars, 1 star, 2 stars, or 3 stars. CPCV does not set that star. A Candidate Strategy is not Alpha.

Common questions

What is combinatorial purged cross-validation?

Combinatorial purged cross-validation splits a time series into blocks and rotates which blocks are held out as the test. Purging removes training observations whose labels overlap a test block, so one split is less able to see the test. The point is to see whether a result depended on a single way of cutting the sample.

Does Alphavo use CPCV to show a strategy?

No. This engine does not use combinatorial purged cross-validation as a gate. CPCV does not set 0 stars, 1 star, 2 stars, or 3 stars.

How does CPCV differ from one train-test split?

One split can flatter a rule that happened to fit that cut. CPCV repeats the check across many combinations of held-out blocks, with overlapping labels purged, so a single lucky cut is harder to mistake for a stable result.