Failure to Improve a Heuristic Multi-Factor Equity Screen: A Negative-Results Study with an Explicit Look-Ahead Audit

We report a negative result and measure the bias that hid it. A five-factor composite equity screen, whose weights were set heuristically and never validated, was tested with standalone rank information coefficients (IC), leave-one-out ablations, a continuous weight sweep, and a volatility-penalty sweep, on 71 overlapping score dates (2020-08 to 2026-07, about 24 effective independent windows) over a 532-ticker US large-cap universe. Three of its five factors were computed from one overwritten, undated fundamentals snapshot, so each stock's 2026 figures ranked it in 2021. We rebuilt quality (return on equity and free-cash-flow yield) and earnings drift (seasonal-random-walk SUE) from SEC EDGAR as they were known on each date and re-ran the tests on identical names. With the snapshot, the screen's top picks beat the average name in their pool by 4.37 points per 63-day window (t_HAC = 3.39; positive on 72% of dates). With dated data, the same screen beats it by -0.65 points (t_HAC = -0.63; 95% interval about -2.8 to +1.5): the difference, -5.01 points per window (t_HAC = -2.79; positive on 27% of dates), is the whole measured edge and slightly more. The look-ahead is one of timing, not restatement: restated figures with a fixed 90-day lag reproduce first-filed ones (rank correlation 0.96-0.97). The finding that removing the earnings-drift factor harms returns (t = -3.2 on the matched names) disappears (t = -0.2), and the penalty on raising the trend weight fades (from t between -2.1 and -2.8 to |t| <= 1.1). The same re-estimation removes the support for the other things we hoped to tune. The selection gates, tuned on the snapshot composite, show a spread of -0.5% on dated data and no setting beats them out of sample; a search over eight pre-specified candidate rankings finds no robust best (the a-priori equal-weight blend returns -1.1%, and the apparent winner, 12-1 momentum, is -2.7% in the first half of the sample and +18.5% in the second and is concentrated in recently added index members); and rank-based Sell labels have no support. No re-weighting improves on the incumbent at a corrected level on either data arm, and with 24 windows the study cannot detect an average IC below about 0.08. We argue that the decision on the weights is inaction, that evidence must now come from a pre-registered, hash-chained forward test of the published screen against shadow arms, and that the contribution is the measured size of the look-ahead bias on a deployed screen.