Does the VIX predict crypto risk? The year it should have worked is the year it scored worst.
Short answer: no — and the year everyone remembers is the year it was most wrong. Scored out-of-sample against a forward drawdown label on bitcoin, the VIX comes out at an AUC of 0.490 over 2,281 days — where 0.50 = coin flip and 1.00 = perfect. Split the same data by calendar year and the average is 0.502: four of ten years above the flip, six below, none of them holding for two years running. And 2022 — the year the VIX and crypto sold off together, the year "watch the VIX" felt most obviously right — scores 0.405, the worst in the sample. This post is about why an indicator can be that visibly correlated with crypto and still not forecast it.
Why anyone asks
The VIX is the options market's estimate of how much the S&P 500 will swing over the next month — Wall Street's fear gauge. Crypto traders watch it because the intuition is compelling: when equities get scared, risk assets sell off together, and crypto is the highest-beta risk asset in the room. "Watch the VIX" is standard advice.
It is also testable, which is the entire point of this series. We already did this to crypto's own sentiment gauge and found the Fear & Greed Index scores a coin flip over eight years. The VIX is the more serious candidate, and it fails in a more instructive way.
How we scored it
Identical method to the last one, so the two are comparable:
- The label. A 10-day forward drawdown on bitcoin in the worst 30% of historical 10-day windows. This is a research screen, not the number the cockpit shows — our headline drawdown read is a much tighter event (a drop from today's price, at its worst point, within 1 day).
- The base rate. 30% of windows qualify by construction, so a useless predictor scores 0.50.
- The direction. A higher VIX counts as the warning, matching our panel and the standard reading.
- The scoring. Combinatorial purged cross-validation with an embargo equal to the horizon. Out-of-sample throughout.
- The sample. 2,281 days of overlapping VIX and bitcoin history, August 2017 to September 2026. Crypto trades every day and the VIX only on trading days, so the count is the days both have.
The result
Pooled across everything: AUC 0.490. Precision-recall 0.302 against a 30% base rate. A coin flip, fractionally on the wrong side.
By year (above 0.50 = a high VIX preceded the rough windows; below = a low one did):
- 2017 (from August) — 0.455
- 2018 — 0.457
- 2019 — 0.514
- 2020 — 0.571
- 2021 — 0.485
- 2022 — 0.405
- 2023 — 0.638
- 2024 — 0.464
- 2025 — 0.548
- 2026 (to 13 September) — 0.481
Average of those: 0.502. Best year: 0.638. Worst: 0.405. The pooled number and the yearly mean agree with each other, and both agree on nothing. There is no year that is followed by another year above the flip; the two readings that look like something — 2020's 0.571 and 2023's 0.638 — each gave way to a below-flip year immediately.
2022: correlated, and wrong
Look at 2022 first, because it is the year the case for the VIX rests on. The VIX spent much of the year elevated; bitcoin lost roughly two-thirds of its value. Put the two on one chart and they move together, unmistakably. And scored as a forecast, within that year, the VIX ranked bitcoin's dangerous ten-day windows worse than a coin flip: the rougher windows more often began on the lower-VIX days.
Those two facts are not in tension. The VIX is a measure of stress now — it rises as the sell-off happens, and it peaks when the damage is mostly done. A forward label asks a different question: from today, are the next ten days going to be unusually bad? By the time the VIX is screaming, the answer has often already been paid for; the days that opened on a high VIX were, as often as not, followed by the relief bounce rather than the next leg down. A coincident gauge and a leading one look identical on a chart and opposite on a forward score.
This is the purest version of the mistake this series keeps finding: co-movement is not prediction. The VIX and crypto risk genuinely move together. That is exactly why it cannot warn you: by the time it moves, so has the thing you wanted warning about.
Why we show the pooled figure and the yearly one
AUC asks a pairwise question: take one risky day and one calm day, and how often does the indicator rank them correctly? Pool ten years together and many of those pairs are cross-year pairs — a day from 2022 against a day from 2021. If the indicator and the risk both drift with the regime, the pooled number can pick up that drift and report it as skill, even when no single year shows any. We have watched that happen: the first version of this post, on a shorter sample, found a pooled 0.537 sitting above every one of the years inside it.
On this longer sample the artefact happens to net out — 0.490 pooled, 0.502 by year — which is the honest way for it to go, but not one you can count on. So we show both, every time. When the two disagree, the yearly figure is the one that answers the question you actually face, because you live through the market one day at a time and never get to trade the difference between two years.
Two different traps, one lesson
It is worth putting this next to the Fear & Greed result, because the failures are not the same shape.
- Fear & Greed runs one way in bear markets and the other way in the year that flatters it; its one strong reading (0.640) is the latest complete year, which is the trap — you would be discovering it from the regime it happens to fit.
- The VIX has no strong regime at all. Pooled or yearly, it is a coin flip, and the year it co-moved with crypto most visibly is the year it forecast worst. The trap here is the chart: a gauge that tracks the risk looks like a gauge that predicts it.
One lesson covers both: a single headline number is not a track record, and a chart is not a score. Ask when the skill was there, not just whether it averages out. That is why every skill figure we publish carries a per-year breakdown beside it — not as a nicety, but because we have watched the pooled figure and the yearly figures disagree in both directions.
What this does not prove
- Not "the VIX is useless." It is a fine description of equity-market stress and it belongs on a macro dashboard. What it does not do, on this label and horizon, is tell you whether the next ten days are unusually dangerous for bitcoin.
- Not "buy when the VIX spikes." Six years score below 0.50, and 2022 is the kind of number that tempts an inverted rule. Our cross-fold standard deviation is 0.045 and each year is roughly 250 trading days, so individual years are noisy, and 2020 and 2023 point the other way. The claim we stand behind is the absence of a consistent edge in either direction — not the direction of any one year.
- One label, one asset, one horizon. A different question might get a different answer, and a monthly horizon in particular is a genuinely different test we have not run here. Our first version scored SOL instead of bitcoin and reached the same verdict by a slightly different route.
- A univariate screen. This measures the VIX alone. Inputs that are weak by themselves can still earn their place inside a multivariate model, and that is not what this number is.
Three questions to ask any indicator
You do not need our data to apply this. Whenever someone shows you a gauge, an index or a signal, these three questions separate a measured claim from a vibe, and almost nothing survives all three.
- "Scored against what, exactly?" Not "does it work" but: what event, over what window, compared to what baseline? If the answer is not a specific event with a base rate attached, there is no claim to evaluate. "The VIX warns you about crypto" is unfalsifiable. "The VIX ranks the worst 30% of 10-day bitcoin windows correctly 49% of the time" is a claim you can be wrong about.
- "In-sample or out-of-sample?" Any indicator can be tuned to look brilliant on the data used to tune it. The number only means something if it was produced on data the method had not seen — which is what cross-validation with a purge and an embargo is for, and why we say so every time.
- "Does it hold in every period, or only on average?" This is the one almost nobody asks. Ask for the breakdown. If the aggregate is strong and no individual period is, the aggregate is measuring regime differences, not forecasting skill — and if the chart looks convincing and the score does not, you are looking at co-movement.
A useful habit: when a number surprises you in a good way, assume it is a pooling artefact or a lookahead until you have ruled both out. That instinct costs you a few hours and saves you from building on a result that was never there. It is also, uncomfortably often, the difference between a backtest and a live drawdown — ranking skill that only exists on a chart does not show up in your account.
Why we keep publishing these
Because the alternative is asking you to take our word for it. Anyone can put a gauge on a dashboard; the question that separates a tool from a decoration is whether its numbers have ever been scored, and scored in a way that could have embarrassed the people doing the scoring.
In w4rn every series shows its measured skill, per year, next to the reading — including the ones that come out looking like this. Our own forecasts are cross-validated and Platt-calibrated, so a stated 30% is meant to be a real 30% — a target we check against what actually happened, not a promise. If that sounds like an odd thing to aim for, that is what calibration means, and it is a much harder bar than "the backtest looked good".
Method note: figures from our univariate forward-AUC screen, combinatorial purged cross-validation, panel dated 13 September 2026, 2,281 observations for the VIX. Offline evidence, not a live trading claim.
Revision note, 14 September 2026: this post originally scored the VIX against SOL over 2020–2026 (0.537 pooled over 1,483 days against a 0.441 yearly mean — the pooling artefact it was written around). It now scores bitcoin over 2017–2026, the asset most readers mean by "crypto", with three more years of history. The verdict is unchanged; the mechanism on show is a different one, and we think a more useful one.
Revision note, 24 September 2026: an earlier version of the paragraph on calibration said that when our forecasts state 30%, those events happen about 30% of the time. That is what calibration aims for, and we had not yet published a check of it against the live record, so the line now states the target rather than a result.
See today's live read in the cockpit → Nothing here is financial advice.