Does the Basket Lead the Fund? Constituent Lead-Lag, Breadth and Overnight Gaps in U.S. ETFs

A popular practitioner idea is that the large constituents of an exchange-traded fund (ETF)—or, for single-country funds whose home market is closed during U.S. hours, the fund’s U.S.-listed American depositary receipts (ADRs)—move first, so that their recent returns predict the fund’s next move. We test this, and three related ideas, on one-minute data for QQQ, six concentrated U.S. sector ETFs and seven single-country ETFs over 2016-06-01 to 2026-07-14. Every result is regenerated from code. Funds track their baskets closely within the same bar (contemporaneous R² of 0.15–0.92). In the cross-correlation scans, the peak is at lag zero for all 14 funds. Lagged basket returns do carry statistically detectable information: 20 of 42 walk-forward Clark–West tests survive a Holm correction. But the median out-of-sample R² is only 0.039%. A one-bar threshold rule never clears realistic costs except where the ETF’s own last-trade price is stale (SOXX in 2018–2022, whose breakeven tracks its missing-minute share). Its apparent profitability at 5 and 15 minutes depends on the arbitrary phase of the bar grid. Two earlier findings do not survive. First, the claim that broad overnight gaps “fade harder” was produced by measuring constituent gaps from a morning-only price file; with official closes the breadth interaction disappears (t = 0.19 vs. −2.99). Second, the claim that the tech overnight return is “uncorrelated with the market” came from a one-day timing misalignment. Correctly aligned, its correlation with SPY is 0.59. Buying tech after down days is the weaker half of a decaying overnight drift. We also document how converting only one of two files stamped in Coordinated Universal Time (UTC) manufactures a spurious four-hour “lead”. Declaration of generative AI use: All research questions, hypotheses, study designs and conclusions are the author’s own, including the decision to re-audit the author’s earlier exploratory results, which identified data errors in them. Under the author’s direction, generative AI tools (Claude Code, Anthropic) wrote most of the analysis code and ran it, and assisted with drafting, editing and formatting the text, and with typesetting. The author reviewed all code, results and text and takes full responsibility for the content of this paper.