Skip to content

Do factor-timing claims beat a placebo?

Four published factor-timing claims, retested with placebos that keep a market's volatility clustering. Volatility-managed factors kept their in-sample verdict under every block length. Industry momentum kept it only with the shortest blocks, and cross-sectional factor momentum lost it. No verdict after the papers' samples changed.

Arhan Canli 3 min read

Published 2026-11-08. The design, the decision rules and the predictions were registered and pushed before the run, in placebo-v2r2-prereg.json; every verdict below is read from placebo-v2r2-evaluation.json. The first round is placebo-v2-evaluation.json.

The question

A factor-timing paper tests a rule, often one of several it could have chosen, on decades of factor returns. A t-test on the winner ignores that search. A placebo test does not: it reruns the whole search, every rule in the grid, on the same returns put in a random order, many times, and asks how often the real winner's result is beaten by chance. This study does that for four claims:

  • Time-series factor momentum (Ehsani and Linnainmaa; Gupta and Kelly): a factor that rose over the past months keeps rising.
  • Cross-sectional factor momentum (Ehsani and Linnainmaa; Arnott, Kalesnik and Linnainmaa): the factors that did best recently beat the ones that did worst.
  • Volatility-managed factors (Moreira and Muir): scaling a factor down when its recent volatility is high raises its Sharpe ratio.
  • Industry momentum (Moskowitz and Grinblatt): the industries that rose most keep rising.

Each claim is tested on the papers' own sample and on the months after it, on Kenneth French's factors and industries. The wording rule was set beforehand: we do not say a published paper is wrong; we report which claims keep a p-value below 5% once a plausible search is counted.

What the second round changed

A plain shuffle breaks every link between one month and the next, including the clustering of volatility that real returns have, so a rule that only exploits clustering can look like timing skill. The second round shuffles whole blocks of 3, 6 or 12 months instead, which keeps short-range clustering, with 999 placebos per test. It also runs White's Reality Check, Hansen's SPA test and Romano and Wolf's StepM on the same grids, measures how much power each method has against timing effects of stated sizes, and repeats factor momentum on the Open Source Asset Pricing long-short portfolios rather than the handful of French-library factors.

What it found

Claim First round, in sample Blocks of 3 months Blocks of 6 Blocks of 12
Volatility-managed factors below 5% below below below
Industry momentum below 5% below not below not below
Cross-sectional factor momentum below 5% not below not below not below
Time-series factor momentum not below not below not below not below
  • Volatility-managed factors are the one in-sample result that survives at every block length.
  • Industry momentum survives only with the shortest blocks, and the evaluation marks that verdict as showing time-series structure rather than the claim itself.
  • Cross-sectional factor momentum loses its first-round verdict at every block length. On the Open Source Asset Pricing factors neither factor-momentum claim beats its placebos in sample.
  • After the papers' samples, no claim's verdict changed: none was below 5% in the first round, and none is now.

What did not work

The registration predicted that block permutation would bring the test's false-positive rate back to its nominal level in simulated markets with clustered volatility. It largely did not, and the evaluation file reports the rates cell by cell. The power runs also show that the tests after the papers' samples have little power against timing effects of the sizes tested, so a claim that is "not below" there is weak evidence against it, not evidence that it is false. One change was made during the run and is dated in the file: four of the 30 industry portfolios are identical to their 12-industry counterparts, so the simulator copies each twin instead of drawing it.

Use it

The placebo test is placebo_test in canli-mcp, on your machine, with no account and no key. The first round's method and its size study are in Would your backtest beat a placebo?.