Purged & embargoed K-fold

mediumMLacceptance —~30 min

Implement purged_folds(n, k, purge, embargo) — K-fold splits for a time series of n observations. Cut the index into k contiguous test blocks (the remainder goes to the last block). For each block, the training set is every index NOT in [test_start - purge, test_end + embargo]. Return one row per fold: [test_start, test_end, train_count] with test_end inclusive.

Random K-fold on a time series lets tomorrow leak into today's training set; purging removes the overlap before the test block and the embargo removes the echo after it. The direction models in the ai quants folder are evaluated exactly this way.

example 1

input

n=100, k=5, purge=2, embargo=0

output

[[0, 19, 80], [20, 39, 78], [40, 59, 78], [60, 79, 78], [80, 99, 78]]

example 2

input

n=100, k=5, purge=2, embargo=3

output

[[0, 19, 77], [20, 39, 75], [40, 59, 75], [60, 79, 75], [80, 99, 78]]

constraints

k ≥ 2
n ≥ 4·k
0 ≤ purge, embargo ≤ n // k

teacher · 2 hints · the first is free

↗ read the theory

the chapter this problem is cut from

▶ open in the workbench

run the same math on a real tape

●solution.py· utf-8 · spaces:4

def purged_folds(n: int, k: int, purge: int, embargo: int) -> list:
    """[test_start, test_end, train_count] per fold; test_end inclusive."""
    size = n // k
    out = []
    for f in range(k):
        start = f * size
        end = n - 1 if f == k - 1 else start + size - 1
        lo = max(0, start - purge)
        # TODO: embargo — drop the `embargo` observations right after the test block too
        hi = end
        train = n - (hi - lo + 1)
        out.append([start, end, train])
    return out

ln 1, col 1tab indent · esc leave editor

python 3.13 · pyodide

+ 6 hidden on submit

input · keyword arguments to purged_folds()

n=100, k=5, purge=2, embargo=0

expected

[[0, 19, 80], [20, 39, 78], [40, 59, 78], [60, 79, 78], [80, 99, 78]]

the judge calls purged_folds() with exactly these inputs · outputs are compared to 6 decimal places

● standing bypython 3.13 in your browser · pyodide 0.28.3 · judged on the server to 6 dp⌘↵ run · ⌘⇧↵ submit

paper only · educational · not advice