Archive / Equities & volatility / E924
Next-day stock reversal in stressed markets did not cover costs
Buying the previous day's biggest large-cap losers and shorting its biggest winners, only on high-volatility days, earned +1.0 bp gross per $1 in 2020-2026, against a 7 bp cost.
Stocks that fell hardest yesterday are said to bounce back, because forced sellers push the closing price too low and market makers, short of capital in stressed markets, are slow to absorb it. We tested this only on high-volatility days, buying the 10% of large caps that fell most and shorting the 10% that rose most, one day, market-hedged. In 2020-2026 the 344 stressed days earned +1.0 bp gross (1 bp is one hundredth of a percent) against a 7 bp cost. The one-week-later placebo earned -5.1 bp.
Large-cap stocks that fell the most today bounce back tomorrow, and the bounce is biggest when markets are under stress.
Why people believe it
The argument is that forced sellers (leveraged funds meeting margin calls, investors redeeming from funds) want to trade at the close and pay for speed. Market makers and statistical-arbitrage desks normally absorb that flow, but in stressed markets their risk limits tighten and they absorb less, so closing prices overshoot and the overshoot is repaid the next day. On calm days the same reversal is said to be too small to notice.
Why it would persist The argument, as stated before testing
The argument is that the scarce resource on stressed days is risk capital itself, so arbitrage capital cannot simply arrive and remove the gap, and the payoff would be compensation for bearing risk, not a pricing mistake.
Each day we measured how much every large US stock beat or lagged the market after removing its normal market sensitivity. We ranked the names, bought the 10% that lagged most and sold short the 10% that beat most, at the next day's opening auction, and closed both at that day's closing auction. We did this separately for stressed days (volatility index at or above its 80th percentile so far), calm days (at or below its 20th) and the rest. A one-week-later placebo and a second, independent universe of sector funds were used as checks. The rule and the pass and kill lines were fixed before any result was computed.
In both 2010-2019 and 2020-2026: stress net mean above 0 with a month-block 95% interval lower bound above 0, stress gross mean at least 21 bp (3 times the cost), and stress gross above calm gross.
Stress gross mean in 2020-2026 at or below 7 bp (the cost). Otherwise neither pass nor kill.
7 bp per round trip per $1 gross: broker commission, auction fees, a short-borrow allowance, the index hedge and slippage against the auction price. For the sector-fund check 3 bp.
Calm-day quintile (negative control); the same ranks held one week later (placebo: the open-to-close return on t+6); the long leg and the short leg reported separately; sector-fund arm (2 lowest against 2 highest of at least 9 funds); stress-episode level statistics (runs of consecutive stressed days as the unit); results split by era.
The 2020-2026 stress gross mean was +1.0 bp, at or below the 7 bp cost, which is the pre-set kill line; the 2010-2019 stress net was -1.0 bp with a lower interval bound below zero.
- In 2020-2026 the 344 stressed days earned +1.0 bp gross and -6.0 bp net of the 7 bp cost (month-block 95% interval for net: -17.1 to +4.6 bp); the median stressed day was -0.6 bp and 49% were positive. Calm days earned +0.0 bp on 105 days, so stressed minus calm was +1.0 bp (interval -12.0 to +12.8 bp).
- In 2010-2019 the 160 stressed days earned +6.0 bp gross, -1.0 bp net (interval -7.2 to +4.8 bp), against -0.9 bp gross on calm days. Stressed minus calm was +6.9 bp, with an interval from -0.2 to +13.6 bp.
- In 2000-2009 the 507 stressed days earned +10.1 bp gross against +3.4 bp on calm days; this is +3.1 bp net of the 7 bp cost, with an interval from -4.1 to +10.2 bp that includes zero; 2000-2009 is not part of the pass test. About 84% of that gross mean sits in the long leg (losers +17.0 bp, winners -3.3 bp), the leg a universe of current index members flatters because failed names are missing.
- Stressed episodes (runs of consecutive stressed days) averaged +15.3 bp in 2000-2009 over 44 episodes (t = 3.10; gross, before the 7 bp cost, in the era whose universe leaves out failed companies, and outside the pass test), -3.5 bp in 2010-2019 over 22 episodes (t = -0.36) and -3.0 bp in 2020-2026 over 43 episodes (t = -0.38).
- The placebo, with the same ranks traded one week later, earned +1.6 bp (2000-2009), +5.4 bp (2010-2019, t = 2.10) and -5.1 bp (2020-2026) on stressed days, so the 2010-2019 stressed-day number was not specific to the day after.
- By volatility quintile in 2020-2026, gross returns were +0.0, +1.3, +0.3, +6.3 and +1.0 bp from calmest to most stressed: the fourth quintile was best at +6.3 bp, still below the 7 bp cost (negative after costs), and the most stressed one was not, so there was no steady gradient with stress. The worst stressed day lost 553 bp.
- The sector-fund arm (3 bp cost, mechanism check only) earned +8.3 bp gross on stressed days in 2000-2009, 0.0 bp in 2010-2019 and -2.7 bp in 2020-2026 (episode mean -10.5 bp, t = -2.16, 35% of episodes positive).
Stressed-day result after costs, by era
Mean net result per stressed day after the 7 bp cost, with month-block 95% intervals. Every interval includes zero, and the two eras that count (2010-2019, 2020-2026) have a mean below zero.
Chart data (bp)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| 2000-2009 (survivorship-flattered) | 3.1 | -4.1 | 10.2 |
| 2010-2019 (160 days) | -1 | -7.2 | 4.8 |
| 2020-2026 (344 days) | -6 | -17.1 | 4.6 |
Stressed days against calm days, gross
Stressed daysCalm days
Gross mean per $1 of gross position, before the 7 bp cost. The pass rule needed stressed days at 21 bp or more in both recent eras; the best era reached 10.1 bp gross, +3.1 bp net with an interval that includes zero (long-leg driven, survivorship-flattered), and 2020-2026 reached 1.0 bp, with calm days at 0.0 bp.
Chart data (bp)
| Stressed days | Calm days | |
|---|---|---|
| 2000-2009 | 10.1 | 3.4 |
| 2010-2019 | 6 | -0.9 |
| 2020-2026 | 1 | 0 |
The next day against one week later
Next dayOne week later (placebo)
Stressed-day gross mean from the same rankings, traded the next day (the test) and one week later (the placebo). A one-day repayment should show up only in the first bar; in 2010-2019 the placebo is nearly as large, and in 2020-2026 the next day is about zero (+1.0) and the placebo is -5.1.
Chart data (bp)
| Next day | One week later (placebo) | |
|---|---|---|
| 2000-2009 | 10.1 | 1.6 |
| 2010-2019 | 6 | 5.4 |
| 2020-2026 | 1 | -5.1 |
Why it failed Interpretation · not established
Over 2000-2009 the stressed-day point estimates point the way the idea says, but they rest on the long leg of a universe that leaves out failed companies, and no era delivers a net gain with an interval above zero. In 2010-2019 the same stock ranks earned almost as much a week later as the next day, which fits a general drift in stressed periods rather than a one-day repayment. In 2020-2026 the stressed-day result shows no detectable difference from the calm-day result. This is interpretation, not established.
Lesson
A reversal that was weak on average was no better on stressed days once tested separately. A state-specific pattern can be driven by a few episodes, one leg of the book or a flattering universe; checking the same ranks a week later, each leg alone and the episode as the unit of independence showed more than the pooled mean did.
Scope of this result
Covers 503 current S&P 500 members, one-day open-to-close holding from auction to auction, market-hedged, a 7 bp cost and a VIX-percentile definition of stress, from 2000 to October 2026. Prices are daily opening and closing prints, not actual auction fills, and the member list reflects today's index, so failed names are absent. The sector-fund arm is a mechanism check, not a traded result.
A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.
1. Test design
Pre-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 5Finding 6Finding 73. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade