VIX Trading Strategy Tested: 2,520 Backtests Say the Roll Pays, Not the Timing

TL;DR: the answer box

• The win rate is real. The money is not. 374 trades, 74.6% winners, 39.62% of the time in the market. Average trade: $173 across twenty years. $1,298 in 2009. $1,267 in 2021.
• Doing nothing beat doing something. Short and hold made $187,070. The timed strategy made $64,720, which is 34.6% of it.
• Holding the combined short 20 days made more money than exiting the same day. Net profit rises from $33,140 to $122,790 and profit factor from 1.06 to 1.45, while the worst drawdown goes from $59,210 to $63,800 and the trade count falls from 1,216 to 186.
• The VIX 20-30 block was profitable in 28 of 28 cells. Median profit factor 1.77. Only 12 of 28 cells with the VIX between 15 and 20 did, median profit factor 0.96.
• Neither version is survivable at one contract. Worst drawdown was $54,620 for the strategy and $77,420 for short-and-hold, against $35,000 of capital. One trade lost $26,060.

How we tested

The instrument is the Cboe VIX futures contract, symbol VX, at $1,000 per index point. A one-point move is a thousand dollars. The data is 4,924 daily bars from 2007-01-04 to 2026-07-02, on the round-the-clock session, which covers 2008, the 2018 volatility blow-up and 2020. Product context lives on Cboe's VIX volatility products page.

Three published entry rules, all short:

1. HH5. Today's high takes out the highest high of the prior five days. Fired on 70.1% of signal days.
2. Up3. Three consecutive higher closes. Fired on 27.7%.
3. RSI2>75. RSI(2) above 75. RSI, or relative strength index, is a standard 0-to-100 gauge of how far price has run in one direction. A 2-period version reacts within days. Fired on 79.4%.

Those are not three ideas. On 19.6% of signal days all three fired at once, and on another 38.2% two of them did. There were 1,216 signal days in total.

Two exits, reported separately and never mixed. The first is the rule these signals are published with, exit when RSI(2) drops below 25 or after 10 bars, whichever comes first. It is simulated directly, because a conditional exit and a fixed hold are different strategies. The second is the engine's sweep across fixed holds of 0, 1, 2, 3, 5, 10 and 20 days. 2,520 variants in total, across both directions, four signals, five VIX regimes, both regime filters and all seven holds. Every fill is the next bar's open. $35,000, one contract, no compounding, frictionless.

One thing about the price file matters more than any of that. A VIX future expires every month, so a 19-year series only exists by stitching contracts together and removing the roll gaps. Against the real Cboe index, this file sits about 192 points too high in 2007 and about 3 points too high in 2026. Point moves and dollar profits are exact on it. Anything divided by the price is not, so percent-of-price figures are excluded from this article rather than reported carefully. Anything that needs a real VIX level, a rule such as "the index was under 15", comes from a second file: the official Cboe VIX daily history, 9,260 days from 1990-01-02 to 2026-08-27, frozen with its provenance.

Finding 1: Do the three published rules hold up?

The win rate is the number that makes these rules look good, and it is real. It is also the number that tells you least.

Four of those seven losing years still won more trades than they lost. 2008 is the one that settles it: a 72.73% win rate and -$36,670. The rules themselves are specified precisely enough to rebuild from, so nothing here turns on a vague definition. What the win rate hides, it hides on its own.

The average trade hides the same thing. Across twenty years it is $173. In 2009 it was $1,298 across 25 trades. In 2021 it was $1,267 across 21 trades. Those are the only two years whose average trade cleared $1,000.

Here is each rule on its own, at the exit they are published with, short side, one contract:

The three published rules and their union, at the exit they are published with.

The long mirror is the last row for a reason. Run the identical entries and the identical exit in the opposite direction and the result is the exact negative. There is no asymmetry hiding in the rules. There is only a direction.

Finding 2: Does the timing beat simply staying short?

No. It is not close.

One contract, the same 4,924 bars, no signal in the top row at all. Long-and-hold here is the arithmetic mirror of short-and-hold. The engine's long buy-and-hold control on the same bars is -$187,170. Different method, same conclusion.

Staying short made 2.9x what the strategy made. No entry rule, no exit rule, no RSI. Short from the first bar of the stitched series to the last, 19 years of contract rolls embedded in the price rather than simulated.

The strategy has two genuine defences and they are worth stating plainly. Its worst drawdown was smaller, $54,620 against $77,420, and it was only exposed 39.62% of the time, so it spent three days in five in cash. Buying a smaller hole and less exposure while keeping 34.6% of the profit is a real trade-off, not a failure. It is just not a trade-off you can see until the two sit side by side, which is what the table above does.

And it is worth being precise about why sitting short works at all. The curve is usually in contango, meaning the next month trades above the front month. This study does not simulate the individual rolls from one contract to the next, or what they cost. It measures their sum: the back-adjusted series embeds 187.07 index points of accumulated roll over these 19.5 years, which at $1,000 a point is $187,070 per contract, and that is short-and-hold's entire result.

Think of a short VIX futures position as a flat you rent out. The rent arrives whether you do anything clever or not. Short-and-hold is the landlord who never leaves. The three rules are the landlord who keeps the place empty three days out of five and is surprised the income is lower.

Finding 3: Where does the profit actually come from?

From holding the combined short 20 days instead of exiting the same day: $122,790 against $33,140, at worst drawdowns of $63,800 and $59,210.

Combined signal, all regimes, both regime filters neutral.

Net profit rises from $33,140 at a same-day exit to $122,790 at twenty days. Profit factor rises with it, 1.06 to 1.45. The worst drawdown goes from $59,210 to $63,800, and the ladder is not smooth: the two-day rung drew down $87,890. Meanwhile the trade count falls from 1,216 to 186. The 20-day variant made $122,790 on 186 trades. The same-day variant made $33,140 on 1,216 trades.

The 20-day hold earned more than the same-day exit. The two-day rung lost $17,970. These spikes had already traded back through their own signal close by the end of day one in 94.1% of the 1,216 events. The same-day exit still earned less in total, though the two variants took different trades: the backtest is flat-only, so one took 1,216 and the other 186. This study does not split those dollars into roll versus other movement. The continuous series embeds the roll rather than simulating it.

Finding 4: When should you short volatility, and when should you not?

When the index is already between 20 and 30. Least often profitable when it sits between 15 and 20.

Every cell below is the short side with both regime filters neutral: four signals times seven holds, 28 cells per regime.

R-expectancy is average profit per trade divided by the average risk taken to get it. Above zero is an edge. 0.260 is a modest one.

Every single cell in the VIX 20-30 block makes money. All four entry rules, all seven holds, no exceptions. The block next door, VIX 15-20, loses in more cells than it wins and its median profit factor is below 1, meaning the losses slightly outweigh the wins.

That is what a stable region looks like: a whole block, not one best row. The block was still found inside the same 2,520-variant sweep it is measured on, and this study holds back no out-of-sample years to confirm it. Pick any rule and any hold inside VIX 20-30 and every cell in this frictionless sweep is profitable. Pick the single highest cell anywhere else and you are relying on it staying the highest.

In this sweep the VIX level on the signal day moves the result far more than the rule you choose. Split the 112 cells inside the four VIX bands by rule and each rule is profitable in 18 to 22 of its 28, at median profit factors of 1.15 to 1.39. Split the same cells by VIX band and the count runs from 12 to 28 of 28, at 0.96 to 1.77. Which is the reverse of leading with the signal and treating the VIX level as a footnote.

The most defensible single configuration in the whole sweep is the combined signal, VIX between 20 and 30, held 20 days, both filters neutral. It is named because it sits in the middle of the block where every neighbour agrees, not because it is the highest number:

That is $177,570 against short-and-hold's $187,070, 94.9% of the money, for 75.2% of the drawdown ($58,250 against $77,420) and 40.94% of the time in market. It made 3.05 dollars of net profit per dollar of worst drawdown, against 2.42 for short-and-hold. That is the one genuinely useful finding here, and no win rate on its own shows it. The hole is smaller, not small: $58,250 is 166.43% of the $35,000 account, and deeper than the combined strategy's $54,620.

Finding 5: Do you need a VIX calendar to make this work?

No. Two supporting facts about the index hold up, and neither was used as an entry rule. This study does not split the timed rules' dollars by cause.

Seasonality, two legs of three. Measured over 37 years of Cboe data, reading "mid-month" as the 15th:

The word doing the work in the usual seasonal story is "consistent". Two legs earn it.

July is the calmest month at a 17.74 mean close and October the most nervous at 21.64. The first two legs are consistent enough to be worth knowing. The third, at 63.9%, is the weakest hit rate of the three. None of the three is a trading rule: no entries, exits or trade results were attached.

The VIX/S&P relationship, measured on returns. Across 8,395 overlapping days:

Correlating returns, not price levels: two levels that drift in opposite directions correlate strongly without telling you anything about a given day.

That is a background fact about the index. It is not an entry. The short-and-hold money on this page is the accumulated roll. A calendar and a correlation do not explain it. This study does not split the timed rules' dollars into roll versus other movement.

The verdict, and the honest limits

The three rules do win about three trades in four. They are in the market about two days in five, and short volatility is the right direction. That part survives contact with 4,924 days of data.

What does not survive is the framing. A 74.6% win rate reads like a discovery, and it is the least informative number on this page. Look at the years:

• 7 of the 20 years lost money.
• Four of those losing years still won more trades than they lost.
• 2008 is the clearest: a 72.73% win rate and -$36,670.
• 2018 went the other way, a 47.62% win rate and -$19,540, which is what a volatility blow-up looks like from the short side.

And the strategy sits inside a market that pays you to hold the short at all. Put the two side by side and the timed rules earned $64,720 against $187,070 for staying short the whole time. Smaller hole, less exposure, less of the money that was already on offer. That is a trade-off, not a discovery. The signals spend 39.62% of the time in the market. That collects less of what staying short already paid.

The one thing that genuinely earns its place is the regime split. The combined signal, VIX already between 20 and 30, held 20 days, both filters neutral: $177,570, a $58,250 drawdown, 40.94% exposure. That is 94.9% of short-and-hold's $187,070 for 75.2% of its $77,420 drawdown, and the drawdown is still 166.43% of the account. Short into a VIX between 15 and 20 and fewer than half the cells made money: 12 of 28, at a median profit factor of 0.96.

Limits

1. Frictionless. No commission, no slippage, no exchange fees anywhere on this page. VIX futures are wider than index futures and the combined rule takes 374 trades, so real costs bite harder here than in most studies.
2. One contract, never scaled. No compounding, no position sizing. Sizing is exactly what a short-volatility book lives or dies on, and this study holds it fixed on purpose so that the signal is what is being measured. The drawdowns are what one contract did, not what a sensible book would have done.
3. The account does not survive any of the three. $54,620 of drawdown for the combined strategy, $58,250 for the VIX 20-30 configuration and $77,420 for short-and-hold, against $35,000 of capital, and a single trade that lost $26,060. Read every dollar figure here as "per contract", not as "what happened to a $35,000 account". On the Mini contract (VXM, $100 a point) they divide by ten.
4. Back-adjusted price. Dollars and points are exact. Percent-of-price metrics are excluded rather than carefully caveated.
5. No roll modelling. The continuous series embeds the roll instead of simulating it. A live trader rolls on a schedule and pays a spread to do it, which this study does not charge.
6. Signals on the close, fills at the next open. These rules are usually written as shorting "when price hits" the level. You cannot act on a close-based signal until the close exists. This rebuild pays that gap. A "when price hits" rule, as written, does not.
7. One market. VIX futures only. None of this transfers to VXX, UVXY or any other volatility product without being retested. Those carry their own decay and their own fees.
8. One thin cell. The VIX-above-30 regime at a 20-day hold rests on 39 trades, below the 50-trade reliability floor. It is quoted nowhere in the conclusions.
9. In-sample only. The VIX 20-30 block and its 20-day configuration were found and measured on the same 2007 to 2026 bars. No years were held back to test them.

Against the seeded random control the picture is honest but unspectacular. Random entries were matched to the median reliable sweep variant: 100 trades, a two-day hold, averaged over 10 seeds. The control made -$5,641.00 on the long basis with a spread of $14,555.67 across seeds, so its short mirror made $5,641.00. The combined rule took 374 trades, so this is not a frequency-matched comparison and the gap is not a clean measure of selection skill. The strategy's $64,720, with its $54,620 worst drawdown, sits well outside that spread. It is still doing less than the contract does on its own.

What this means for you

1. Never read a win rate without the benchmark next to it. 74.6% sounds like an edge. Against a market that pays you to hold the position, it is a way of collecting less of what was already on offer.
2. Ask what the position earns while you do nothing. If the answer is "quite a lot", your entry rule has to beat that, not zero. On VIX futures the do-nothing number over 19 years was $187,070.
3. If you short VIX futures on these rules, short them when the index is already between 20 and 30. That block was profitable in 28 of 28 cells in this 2007 to 2026 sample. The VIX 15-20 block was profitable in 12 of 28.
4. Size for the single worst trade, not the average one. The average trade was $173. One trade lost $26,060. Those two numbers describe the same strategy.
5. On the Mini, every dollar figure on this page divides by ten. VXM is $100 a point against VX's $1,000. This study sizes one contract and does not pick which contract you should trade.
6. A twenty-year average can hide fat years. I lost about $270,000 early in my career trusting numbers that flattered me. Across twenty years the average trade here is $173. 2009 was $1,298. 2021 was $1,267. Those are the only two years whose average trade cleared $1,000. The mistake gets smaller. It does not go away.

Methodology and honesty footer

Cboe VIX futures (VX), daily bars, round-the-clock session, as a difference back-adjusted continuous contract. 4,924 bars, 2007-01-04 to 2026-07-02. Contract multiplier $1,000 per index point. Level-sensitive measurements come from the official Cboe VIX index daily history, 9,260 days from 1990-01-02 to 2026-08-27, frozen with its provenance. $35,000 starting capital, one contract, no compounding, flat-only. No commission, no slippage, no exchange fees anywhere on this page. Signals are read at the close and filled at the next open. 1,272 of the 2,520 variants clear the 50-trade reliability floor. The rest are flagged in the grid and never dropped. Every number here traces to the study's computed results.

This is not investment advice. Past results do not promise future ones. A backtest is a measurement of what happened, not a forecast.

In this 2007-2026 sample, sitting short collected $187,070 of roll. Everything on this page is a decision about how often to leave.

If you want the version of this where I check the benchmark before I believe the win rate, that is the whole newsletter. StatOasis.com/Overfit