Parvanta asks you to design the train, validation and test split for eight years of daily data with 20-day forward-return labels. How many days must be purged and embargoed around each boundary, and how many walk-forward folds with one-year test windows remain?

Purge the 20 training days before each boundary, because their 20-day labels reach into the next block, and embargo 20 days after a test block before any later data is used to train. With three years of warm-up, eight years leave 5 walk-forward folds, testing years 4 to 8. Each fold trains on everything before it, purges 20 days, validates on half a year, purges 20 more and tests a year, losing about 2% of its days to purges.

Suppose a teacher sets a test on chapters 5 to 8 but the last homework sheet, set the week before, covered chapters 4 to 6. A student graded on that homework has already practised part of the test. A 20-day forward label on day t is built from prices on days t+1 to t+20, so a training row in the last 20 days before the test block carries test-period returns inside its label. Train on those rows and the model learns from the very moves it will be scored on. The fix is to purge them: drop every training row whose label window reaches past the boundary. Strictly 19 rows overlap and the 20th ends on the boundary price; purging 20 is the clean rule and costs one day.

The purge handles labels that look forward into the test block. The embargo handles the other direction: rows just after a test block, if they are ever used for training, have features built from test-period prices and labels correlated with the test labels, so a gap of about one label length, 20 days, is left after each test block. In a pure walk-forward design, the model never trains on data after the test block, so the embargo only matters when later data is reused, as in cross-validation that rotates the test block through the middle of the sample, or when the final model is refitted on everything. If features use long look-back windows, for example a 60-day momentum, the embargo should grow with them.

With 2,016 days and a three-year warm-up, 1,260 days remain, exactly 5 one-year test windows. Fold 1 trains on years 1 to 3 and tests year 4; fold 5 trains on years 1 to 7 and tests year 8. Inside each fold, the half-year before the test block is the validation set used to tune settings, with a 20-day purge on each side: train, purge 20, validate 106 days, purge 20, test 252. Each fold gives up 40 days, about 2% of the sample, and the test years never touch the data used to tune the model they score.

Overlap shrinks the information as well as leaking it. A year of 20-day labels holds only about 12.6 non-overlapping periods, so each test year is a small sample even though it has 252 rows per stock. Standard errors that treat every day as independent will be far too small; compute them on non-overlapping labels or adjust for the overlap. Also keep the universe as it stood on each date, including companies later delisted, and make sure every feature uses data available at that day's close. The split is only as clean as the timestamps behind it.