Archive / Equities & volatility / E888
Buying the S&P 500 after a VIX spike rests on three crisis days
Buying the index at the close of a VIX-spike day earns +9.4 bp on average in the 2020-2026 holdout, but three days produce 156% of that profit and the pre-1990 version of the idea loses money.
A popular rule of thumb says to buy stocks when fear spikes. We bought the S&P 500 at the close on days when the VIX jumped more than on 90% of the previous year's days, and sold at the next close. The average looks positive in every era: +12.8 bp, +11.3 bp and +9.4 bp in the 2020-2026 holdout. But in the holdout the three best days made 156% of the profit; without March 2020 and the three best remaining days the result is -4.4 bp. Before the VIX existed, the same idea triggered by sharp index drops lost 27.8 bp per trade.
Buy when fear spikes: after a jump in the VIX, the stock market tends to bounce the next day.
Why people believe it
Volatility-targeting funds, risk-parity portfolios, trend followers and options dealers who are short gamma have to cut equity exposure when volatility jumps, much of it into the closing auction. Whoever buys from them is providing liquidity at a bad moment and should be paid by a rebound the next day. An academic study (2012) of liquidity provision finds that its returns rise when the VIX is high, though for short-term reversal across stocks rather than for the index itself.
Why it would persist The argument, as stated before testing
It is not an arbitrage but a payment for bearing risk at the worst moment: the buyer holds index exposure for a full day while volatility is high.
We used daily closes of the S&P 500 and the VIX. A panic day is a day when the VIX rose more than on 90% of the previous 252 trading days and closed at 15 or higher. On each panic day we bought the index at the close and sold at the next close, after a 1.2 bp cost. The rule was chosen on 1990-2009, checked on 2010-2019 and judged on 2020-2026. Because the VIX starts in 1990, an independent check on 1975-1989 used sharp index drops as the trigger instead.
Holdout net mean at least +8 bp, 95% interval above 0, median above 0, alpha above 0, the same sign in 1990-2009, 2010-2019 and the 1975-1989 control, and the result survives the robustness drops.
Holdout net at or below 0; or median at or below 0 with sign-test p above 0.1; or more than 50% of the profit from 3 days or fewer; or a worst trade below -8% without positive tail compensation; or a control of opposite sign.
1.2 bp round trip; stress case 4.5 bp.
Alpha against the unconditional next-day mean in the same window. Independent-era control 1975-1989: event = S&P 500 one-day return at or below -1.5 trailing 60-day standard deviations, which had to show the same sign. Robustness: drop 2008-09, March 2020 and the best 3 days. Week-block bootstrap intervals; sign test.
The three best days produced more than 50% of the holdout profit (156%), and the independent 1975-1989 control had the opposite sign; these are the two reasons for the kill.
- The average was positive in all three eras: +12.8 bp in 1990-2009 (n 371), +11.3 bp in 2010-2019 (n 193) and +9.4 bp in the 2020-2026 holdout (n 169). All three 95% intervals include zero.
- The typical trade faded: the median fell from +17.8 bp and +11.6 bp to +3.0 bp in the holdout, and the sign test went from p 0.02 in 1990-2009 to p 0.76.
- In the holdout the three best days produced 156% of the total profit, against 33% in 1990-2009 and 65% in 2010-2019. Above 100%, the remaining 166 trades lost money combined.
- Dropping the 5 panic days of March 2020 and then the three best remaining days (n 161) turns the holdout result to -4.4 bp (interval -22.2 to +13.7 bp).
- Before the VIX existed, buying after sharp index drops (1975-1989, n 232) lost 27.8 bp per trade (interval -56.5 to -5.8 bp), and 19.5 bp without October-November 1987.
- Neither alternative trigger held up in the holdout: VIX above VIX3M gave +7.9 bp (interval -41.4 to +61.4 bp) and the 2-day VIX jump +11.0 bp (interval -12.9 to +36.8 bp). After Holm correction the holdout p-value is 1 for all three tests.
- The worst holdout trade lost 952 bp, or 9.5% from one close to the next. That is beyond the -8% worst-trade line of the kill criteria; the verdict does not rest on that line, because the large winning days count as the positive tail that line allows for.
Positive on average, never clearly above zero
Mean next-day net return per panic day, with week-block bootstrap 95% intervals. Every era is positive and every era's interval includes zero. Removing a handful of holdout days flips the sign, and the pre-1990 control is clearly negative.
Chart data (bp)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Selection 1990-2009 (n 371) | 12.8 | -1.3 | 26.4 |
| Validation 2010-2019 (n 193) | 11.3 | -6.5 | 30.7 |
| Holdout 2020-2026 (n 169) | 9.4 | -13.4 | 33.5 |
| Holdout without March 2020 and the 3 best remaining days | -4.4 | -22.2 | 13.7 |
| Control 1975-1989: buy after sharp index drops | -27.8 | -56.5 | -5.8 |
The typical panic day stopped paying
MeanMedian
Mean and median next-day net return per era. The holdout mean still looks acceptable; the median shows that the typical panic day now earns +3.0 bp.
Chart data (bp)
| Mean | Median | |
|---|---|---|
| Selection 1990-2009 | 12.8 | 17.8 |
| Validation 2010-2019 | 11.3 | 11.6 |
| Holdout 2020-2026 | 9.4 | 3 |
Why it failed Interpretation · not established
The numbers fit a conditional equity-risk premium of about +10 bp that, in the holdout, is collected mainly in V-shaped crashes such as March 2020, when a few very large rebounds follow very large falls; on an ordinary holdout panic day the rebound is close to zero. In 1990-2009 the gain was broader (the three best days made 33% of it). The opposite sign before 1990, with a price trigger instead of the VIX, is a separate problem: the idea does not replicate in an independent era.
Lesson
A positive mean in every period is not replication when the median, the sign test and the share of profit from the best days move the other way. For a rule that trades rare stress days, the decisive checks are what remains after removing the few largest days, and whether an independent era agrees in sign.
Scope of this result
Covers the S&P 500 cash index, close to close, a one-day hold after VIX spikes, 1990 to 2026 (control 1975-1989), at a 1.2 bp cost. It does not cover intraday entry, other equity indices or holding periods other than one day.
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