Twelve Rejections: Pre-Registered Kill Criteria and the Non-Monotone Cost of Clean Data in Retail-Scale Systematic Trading

Abstract

Quantitative finance publishes what worked. This paper reports thirteen registered systematic strategy candidates that did not — twelve rejections and one partial — each assessed against kill criteria fixed before the data was examined, on provenance-clean data at retail scale. Five were gated on a t-statistic and a deflated Sharpe ratio and carry the paper's statistical weight; the earlier eight were gated on Sharpe lift and are disclosed to keep the denominator honest rather than to support the argument. The methodological result is what happened when data quality improved: corrections did not move results in a consistent direction. Split adjustment drove post-earnings-announcement drift from t = 0.495 to t = -0.258 (a sign flip), but drove market-residual momentum from t = 1.868 to t = 2.017 — across its pre-registered kill threshold of 2.0. Correcting a separate 2x transaction-cost undercharge then returned it to t = 1.909. For one revision, therefore, a factor appeared to clear the primary significance gate. It was rejected anyway, by a second pre-registered gate — a deflated Sharpe ratio accounting for multiple testing; the recorded verdict at that revision is DEAD_DSR. A single-gate pre-registration would have produced a false positive that survived until the cost bug was found, on a candidate flattered by two defects at once (an uncorrected survivorship-affected universe and an undercharged cost model). We argue that (i) negative results with pre-registered gates are worth publishing, (ii) the common assumption that cleaning data can only reduce apparent edge is false — one counterexample settles it, though how often corrections inflate is unmeasured here — and (iii) pre-registration should specify more than one independent gate, because the gate that catches a marginal result is not reliably the one you expect. Limitations. These are our implementations at our scale; a rejection is evidence about this construction of a factor, not proof the factor is absent from the literature's samples. The equity universe is not point-in-time, so results are survivorship-affected — which makes the rejections conservative but means the single near-miss cannot be read as a near-discovery. Statistics computed before 2026-07-15 predate the cost fix and are marked as such. The denominator counts formally registered experiments only.