Overnight versus intraday returns in the 400 most-traded optionable US names
Across the single stocks in the 400 most-traded optionable US names, does buying at each day's close and selling at the next open beat holding from the open to the close, and does any slice still make money after paying a round trip every night?
The answer depends on what a nightly round trip costs. At the spread-based cost tiers used in the first run, 0 of the 21 statistics is EDGE, 4 are NO EDGE and 17 UNDERPOWERED. The top 100 stocks make +11.21 bps a night gross in the 545-night test window but pay 7.95 in costs, leaving +3.26 (t 0.76). The auction addendum measured a small market-on-close / market-on-open round trip at about 0.11 bps a side in fees, and the auction result file re-scores every statistic at 0.15 bps a side. At that cost 6 of the 21 are EDGE by the study's per-cell rule: the equal-weight 324 stocks (+7.37 net, t 2.21), the top 100 (+10.91, t 2.54), Financials (+10.64, t 2.25), Industrials (+8.17, t 2.15), Information Technology (+12.11, t 2.16) and Utilities (+9.48, t 3.48). Only Utilities clears the 21-test family correction (p 0.0003) and it still does at the 1.0 bps stress (+7.78, t 2.86, p 0.0021); the top 100 also stays EDGE at 1.0 (+9.21, t 2.14) but does not clear the family correction. The auction write-up's answer paragraph names only the top 100 and QQQ, but the corrected write-up now reports all 21 at that cost. None of this is tradeable evidence: the 2026 name list is applied back to 2016 and biases the gross returns up, the cost is fees only at the official print with no spread or impact, the 545 test nights were read rather than held out, and the study's forward paper record of the top-100 book was stopped on fill quality.
Published as a description, not a pre-registered pass or fail: the rules were self-declared before their runs but never pre-registered or approved, and the record was written after the results. At the spread-based cost tiers 0 of the 21 statistics is EDGE. At the measured auction cost of 0.15 bps a side, 6 of the 21 are EDGE by the study's per-cell rule and only Utilities clears the 21-test family correction (p 0.0003; p 0.0021 at the 1.0 bps stress). The auction write-up's answer paragraph names only the top 100 and QQQ, but the corrected write-up now reports all 21 at that cost. The auction cost is fees only at the official print, and the study's forward paper record of the top-100 book was stopped on fill quality.
Backtested by QS Research
Pre-registration
This study was never pre-registered. The two rule documents were written and committed by the same session that ran the work, before each run but with no review or approval: the main cells on 12 November 2026, and the auction-cost addendum later the same day, after the first results were known. The record now with the study is retrospective, written on 21 November 2026 after every result. It records the rules as run: 2,729 nights from 4 January 2016 to 9 November 2026; 324 single stocks from the 400-name list, with 76 funds excluded except as a side note; a train/test split at 9 September 2024 (545 test nights); 21 statistics in one family; and an EDGE only if the test net mean has t of 2 or more in the declared direction and the train net mean is above zero. The addendum fixed the auction cost at 0.15 bps a side by rule (the p90 of fees plus any adverse print deviation, rounded up); its declared re-judging covered the top 100 and QQQ, the auction result file scores all 21, and the corrected write-up now reports all 21. One sensitivity run that drops the 2016 warm-up nights was added after the first results; it is reported, not a cell.
Registered 2026-11-21Amendments
- 2026-11-12 Added an auction-cost measurement and a re-score at a measured market-on-close / market-on-open cost of 0.15 bps a side, declared after the first cost-tier results were known and before any auction measurement. The measurement covers January 2025 to November 2026 on the top-100 names and four index funds; the re-score is the same 21 statistics with only the cost changed. approved: Self-declared 2026-11-12 after the first results were seen; no approval was sought or given.
Trials · 0 trials · 0 passed · judged at N = 21 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| C1 · EW 324 stocks overnight | RESULT | not counted |
| C2 · Top 100 overnight | RESULT | not counted |
| C2 · Ranks 101-324 overnight | RESULT | not counted |
| C3 · Communication Services | RESULT | not counted |
| C3 · Consumer Discretionary | RESULT | not counted |
| C3 · Consumer Staples | FAIL | not counted |
| C3 · Energy | RESULT | not counted |
| C3 · Financials | RESULT | not counted |
| C3 · Health Care | FAIL | not counted |
| C3 · Industrials | RESULT | not counted |
| C3 · Information Technology | RESULT | not counted |
| C3 · Materials | RESULT | not counted |
| C3 · Real Estate | FAIL | not counted |
| C3 · Utilities | RESULT | not counted |
| C4 · Intraday-reversal decile L/S, per $ gross | FAIL | not counted |
| C5 · 20-day momentum decile L/S, per $ gross | RESULT | not counted |
| C6 · Persistence Q5-Q1 (train-ranked), per $ gross | RESULT | not counted |
| C7 · SPY overnight | RESULT | not counted |
| C7 · QQQ overnight | RESULT | not counted |
| C7 · IWM overnight | RESULT | not counted |
| C7 · DIA overnight (bars from 2021-10-25) | RESULT | not counted |
| C2 top 100 overnight · tiers (first run) | RESULT | not counted |
| C2 top 100 overnight · auction 0.15 | RESULT | not counted |
| C2 top 100 overnight · stress 1.0 | RESULT | not counted |
| C7 QQQ overnight · tiers (first run) | RESULT | not counted |
| C7 QQQ overnight · auction 0.15 | RESULT | not counted |
| C7 QQQ overnight · stress 1.0 | RESULT | not counted |
| Close print vs 15:59 minute close | RESULT | not counted |
| Close print vs 15:59 minute VWAP | RESULT | not counted |
| 15:59 minute range (high-low) | RESULT | not counted |
| Open print vs 09:30 minute VWAP | RESULT | not counted |
| Open print vs 09:30 minute open | RESULT | not counted |
| Open print vs last pre-open trade | RESULT | not counted |
| Overnight from prints minus from VWAPs (per night, day-clustered) | RESULT | not counted |
| Fees per side at today's rates | RESULT | not counted |
| C1 · stocks overnight | RESULT | not counted |
| C2 · top 100 overnight | RESULT | not counted |
| C2 · ranks 101+ overnight | RESULT | not counted |
| C3 · Communication Services overnight | RESULT | not counted |
| C3 · Consumer Discretionary overnight | RESULT | not counted |
| C3 · Consumer Staples overnight | RESULT | not counted |
| C3 · Energy overnight | RESULT | not counted |
| C3 · Financials overnight | RESULT | not counted |
| C3 · Health Care overnight | RESULT | not counted |
| C3 · Industrials overnight | RESULT | not counted |
| C3 · Information Technology overnight | RESULT | not counted |
| C3 · Materials overnight | RESULT | not counted |
| C3 · Real Estate overnight | RESULT | not counted |
| C3 · Utilities overnight | RESULT | not counted |
| C4 · intraday-reversal L/S overnight | RESULT | not counted |
| C5 · 20d momentum L/S overnight | RESULT | not counted |
| C6 · persistence Q5-Q1 overnight | RESULT | not counted |
| C7 · SPY overnight | RESULT | not counted |
| C7 · QQQ overnight | RESULT | not counted |
| C7 · IWM overnight | RESULT | not counted |
| C7 · DIA overnight | RESULT | not counted |
Limitations
- Not pre-registered: the rules were written and committed by the session that ran them, before each run but with no review or approval, and the record was written on 21 November 2026 after every result. No result here is a pre-registered test.
- The 2026 list of the 400 most-traded names is applied back to 2016: look-ahead and survivorship. Names that became liquid by rising bias the gross returns up, most in the 545-night test window, and both the six auction-cost EDGEs and the Utilities family-level result rest on this list.
- Price-only: the data has no dividends, so all of a dividend falls on the overnight leg and biases it down by about 0.5 bps a night on average. No verdict depends on it.
- The auction cost is fees only (the SEC Section 31 fee and the FINRA trading fee, no commission), measured on the top-100 names over 2025-2026 and applied flat to all 21 statistics in the auction result file. It assumes a fill at the official print with no spread, no impact, no imbalance and no queue effect. The auction write-up's answer re-judges only the top 100 and QQQ at it, although the corrected table now reports all 21.
- The 0.15 bps cost was measured on 2025-2026 name-days; the 2016-2024 period had a higher SEC rate (about 0.26 bps a side at the peak), and no dated fee model was applied to the backtest. The 1.0 bps stress is reported but is not a verdict: the top 100 and Utilities both remain EDGE at it, and Utilities still clears the family correction.
- A forward paper record of the top-100 book, opened from this study, was stopped on 20 November 2026 by its own kill rule: 238 paper fills were a median 3.76 bps a side from the official prints against a 1.0 bps limit. That measures the paper broker's simulated auction path, not live fills, but the at-print cost was not confirmed forward.
- The auction re-score is a second look: the cost was measured and applied after the tier-cost results were known, and the addendum's declared re-judging covered only two cells; the auction result file scores all 21 and the corrected write-up now reports all 21. The 21-test family correction covers the 21 statistics, not the extra choice of cost model, so the six per-cell EDGEs are not independent evidence.
- The persistence cell has no train-period net: its portfolio is formed from train-period ranks and exists only after the split, so the 'train net mean above zero' leg of the EDGE rule cannot be met for it. At 0.15 bps its test net is +5.99 (t 2.56, p 0.0053) but it is reported UNDERPOWERED.
- The result files record no gross-side standard error, so the gross t-statistics in the write-up (3.0, 3.6 and 2.7) are not in the result files and are only approximate. The published numbers are the recorded test net means and t.
- No holdout was designated: every session in the pinned data, including the 545 test nights, was read. The only untouched data is after the pins' last session, which is not a held-out sample.
- The run's binary commit is not recorded, and the report carries no data snapshot ids even though the result files name the daily bars and the panel; the published provenance therefore identifies neither the build nor the snapshots.
- The post-hoc top-20 names are selected on the test period and are not evidence; 7 of the 20 have a test t of 2 or more. Slices by year, sector and liquidity are descriptive, not tested.
- Short books pay no borrow charge, and the long-short cells are returns per dollar of gross exposure, not investable portfolio returns.
- The Utilities result is 14 names, and its test-period net is concentrated in 2024-2026 (about +11.4, +6.8 and +10.9 bps a night) against +2.08 in train; it is a narrow, recent sector result, not a decade-long premium.
- The sector buckets use each name's current (2026) industry classification for the whole history, a small look-ahead.
- The page's trial list is every row of the write-up tables, including eight measurement rows and the same 21 statistics at the tier and auction costs; all are marked uncounted. Four tier rows carry the write-up's NO EDGE verdict (published as fail); none is counted, and none is a pass/fail result.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current b57ff924277f0e0f, published 2026-11-21