Split every trading day in two: the overnight leg (yesterday's close to this morning's open) and the intraday leg (open to close). In 27 years of QQQ history, essentially all of the return came from the overnight leg — $1 compounded overnight-only grew to about $35 while intraday-only fell to 47 cents. Firms turned this into ETFs. The ETFs died. We froze the claim and are grading it forward.
Every session's two legs are graded automatically at the close. A verdict needs at least 30 post-registration sessions and at least 20 ISO weeks — and the verdict is computed on a single pre-declared date, 4 January 2027, and then locked. Until then the report shows the running numbers and no verdict at all — see the amendment note below. Running numbers live in the calendar experiments report.
We are testing where returns happen, not whether you can harvest them. Capturing the overnight leg costs one round trip per session — 250+ a year — so even small trading costs consume an effect of this size, before slippage and taxes. This isn't hypothetical: the NightShares "night effect" ETFs launched in 2022 to sell exactly this pattern and shut down in July 2023. If this experiment passes, the honest headline is "QQQ's returns accrue overnight," and nothing more. That scope is registered and binding, the same way our risk-score finding is bound to "volatility, not alpha."
Because it is the strongest single structure in our 27-year exploratory sweep — the overnight leg beat the intraday leg in every five-year block since 1999, in the S&P 500 as well as QQQ — and because it is a claim that was literally packaged and sold to retail investors. A structural fact this large, this persistent, and this monetised deserves a public forward test with the prediction written down first. It also has the cleanest control we have ever run: each day's intraday leg is the comparison for that same day's overnight leg. Same day, same asset, same news — the market's open hours are the only difference.
1. At least 30 post-registration sessions graded, and at least 20 distinct ISO weeks.
2. The mean and the median of the per-session difference (overnight minus intraday) both positive.
3. A 95% confidence interval on that difference that excludes zero, computed treating each week as the unit of evidence.
4. The direction holding across at least three consecutive weekly snapshots.
5. All of the above assessed once, on the pre-declared verdict date of 4 January 2027 — then written down and never recomputed.
Win rate is reported but is explicitly not a pass criterion.
The direction has 27 years of history behind it, but the effect is roughly five basis points a day against a hundred basis points of daily noise — at 30 sessions, statistical power is the risk, not direction. Our registered expectation: the difference points positive but has maybe a one in three chance of clearing the full bar at the first snapshot, rising if it's allowed to accrue. And the gap has been narrowing: the intraday leg has been positive since 2010, so a fading effect is a live possibility and would itself be worth publishing.
The interesting outcome isn't only "real or not." It's watching, in public, whether a famous effect that killed two ETFs still exists at all once someone writes the prediction down first.
An adversarial review of our own grading code, run the day after registration, found that the verdict was being recomputed at every look with no correction for repeated testing — so across a year of snapshots a claim with no real effect had roughly a one-in-five chance of printing a pass at least once. Thirty sessions is also only about six weeks, far too few independent weeks for the interval to be trustworthy.
So this experiment now also requires at least 20 distinct ISO weeks, and the verdict is computed once, on a single pre-declared date — 4 January 2027 — then written to a file and never recomputed, so re-running our own code on later data cannot turn a null into a pass. The first read moves from September 2026 to 4 January 2027. Nothing else changed: the claim, the control, the scope limit and the registration date are untouched, and not one row of data had been graded when this was written.
The full diff is in HYPOTHESES.md and in the git history. We are publishing it because a project that grades other people's claims does not get to quietly patch its own scoreboard.
Every session's two legs are appended to experiments/EXP0405-CAL-outcomes.csv — public and append-only. The frozen claim lives in HYPOTHESES.md as H-EXP05, dated, with the constants in calendar_eval.py. Running numbers regenerate into the calendar experiments report; the verdict goes to the audit log and this page. If we change any constant, the test is void and we'll say so.
It is not a tradeable edge, not a recommendation, and not advice — see the binding scope note above. It is not a claim about any fund or firm; the NightShares closure is cited as public context for why the scope stops at attribution. The verdict describes QQQ over this window, nothing broader.
A verdict lands every few weeks. We'll email you each one, free, the day it publishes — and nothing else. No signals, no picks, no offers. Verdicts stay free and public for everyone either way; this just means you don't have to remember to check back.
One confirmation email first — you're not on the list until you click it. Privacy
Educational and informational only — not investment advice, not a recommendation, and not a broker-dealer. ThePickLog is operated by AMD Ventures, LLC (Florida).
← Experiment 04: the turn-of-month effect
Disclaimer · Privacy · All experiments