Does a strong dollar predict crypto drawdowns? Four years say yes, six say no.
Short answer: sometimes, and you cannot tell in advance which kind of year you are in. Scored out-of-sample against a forward drawdown label on bitcoin, the US Dollar Index lands at an AUC of 0.451 over 2,281 days — where 0.50 = coin flip and 1.00 = perfect. Split by year it swings from 0.381 to 0.616, clearing a coin flip in four of ten. Three of those — 2019, 2020 and 2025 — are readings you would happily build on. And the year in between, when the dollar made a twenty-year high and bitcoin fell by roughly two-thirds, the same indicator scored 0.381: the dollar and crypto did exactly what the story says, and the dollar still could not tell you which ten days would hurt.
The claim being tested
"Watch the dollar" is close to consensus in crypto macro. The reasoning is sound on its face: crypto is priced in dollars, so a stronger dollar is a mechanical headwind, and the two do tend to move in opposite directions. Our own macro backdrop carries the DXY for exactly that reason and describes it as an inverse relationship, which it is.
But "moves opposite to crypto" and "tells you crypto is about to get dangerous" are different claims, and only the second one is useful in advance. This is the distinction the whole series keeps circling: co-movement is not prediction. So we tested the second claim.
How we scored it
Same method as the Fear & Greed and VIX posts, so all three are directly comparable:
- The label. A 10-day forward drawdown on bitcoin in the worst 30% of historical 10-day windows. 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 stronger dollar counts as the warning, matching our panel and the consensus reading.
- The scoring. Combinatorial purged cross-validation, embargo equal to the horizon, out-of-sample throughout.
- The sample. 2,281 trading days, August 2017 to September 2026 — long enough to hold two full cycles, which matters for what follows.
The result
Pooled: AUC 0.451. Precision-recall 0.285 against a 30% base rate. Below the coin flip — read literally, the stronger-dollar days were the calmer ones, slightly more often than not.
By year (above 0.50 = a stronger dollar preceded the rough windows; below = a weaker one did):
- 2017 (from August) — 0.459
- 2018 — 0.418
- 2019 — 0.606
- 2020 — 0.611
- 2021 — 0.400
- 2022 — 0.381
- 2023 — 0.434
- 2024 — 0.513
- 2025 — 0.616
- 2026 (to 13 September) — 0.407
Mean of the years: 0.485. Best: 0.616. Worst: 0.381. The spread from best to worst is 0.235 — on an axis where 0.60 counts as a genuinely useful signal and 0.40 counts as a badly wrong one, that is the indicator changing its mind completely, and doing it more than once.
The third way an indicator fails
Three posts in, we now have three distinct failure shapes, and this one is the most dangerous:
- Fear & Greed — a coin flip whose one strong year is the latest one. Trap: discovering an indicator from the regime that happens to flatter it.
- The VIX — a coin flip pooled and by year, and its worst year is the one it co-moved with crypto most visibly. Trap: the chart looks like a forecast.
- The dollar — genuinely good in some years, genuinely bad in others, averaging to nothing. Trap: it switches on and off, and nothing on the chart tells you which state it is in.
Read the years as a sequence. Off in 2018. On for two years, 2019 and 2020, both above 0.60 — a real edge, held for two consecutive years. Then off for three, 2021 to 2023, and not gently: 0.400, 0.381, 0.434. A marginal 2024. On again in 2025, at 0.616, the best reading in the sample. Off again so far in 2026. Anyone who screened this at the end of 2020 would have concluded the dollar works, and been badly wrong for three years. Anyone who screened it at the end of 2023 would have concluded it does not, and missed 2025. Anyone who screened it at the end of 2025 — the last time it looked good — is the reader we are most worried about.
An indicator that is reliably useless is easy to ignore. An indicator that worked, recently, and then stopped is the one that gets built into a position — and this one has done that twice.
2022, specifically
The year deserves its own paragraph because it is the year the consensus points to. The dollar index reached its highest level in two decades; bitcoin lost roughly two-thirds of its value. The inverse relationship the macro panel describes was on full display, and within that year the DXY scored 0.381 — the rougher ten-day windows more often opened on a weaker dollar. Both things are true at once: the dollar and bitcoin moved against each other over the year, and the dollar's level on any given day told you nothing useful about the next ten. The co-movement was real. The forecast was not.
Why it flips
We are speculating past the data here and will say so plainly, but the shape is not mysterious. The dollar is not one signal; it is whatever is driving the dollar at the time. In some stretches dollar strength tracks rate expectations and risk appetite, and those genuinely do lead crypto risk — that is what a year like 2025 looks like. In others the dollar moves on trade flows, foreign politics, or safe-haven demand triggered by a shock that has already hit crypto — at which point the dollar is a lagging symptom of the same event, not a leading indicator of it, and the sign runs the other way.
The composite nature is the problem. When an input's meaning changes with the regime, its measured skill changes too, and averaging across regimes tells you about neither.
What would actually make the dollar useful
If the problem is that the DXY means different things in different regimes, the fix is not a better threshold on the DXY. It is conditioning — knowing which regime you are in before you read the indicator.
Concretely: dollar strength driven by rising real rates is a different animal from dollar strength driven by a flight to safety, and a model that also sees rates, credit and cross-asset volatility can tell those apart in a way that a single line on a chart cannot. That is the honest argument for a multivariate model over a dashboard of gauges — not that the gauges are wrong, but that each one is ambiguous alone and the ambiguity is resolvable with context.
It is also why "this scored 0.451 univariate" is emphatically not the same as "this is useless in the model". A feature can carry real information that only becomes legible next to other features. We publish the univariate number because it is the honest answer to the question people actually ask — can I watch the dollar and know when to be careful? — and the answer to that specific question is no.
The practical version, if you take one thing from this: an indicator whose meaning changes with the regime needs a regime read before it can be used at all. If you cannot state which regime you are in and why this indicator should work in it, you are not using a signal, you are using a habit.
Does this contradict our own macro panel?
No, and it is worth being explicit rather than letting you wonder. The panel says the dollar and crypto usually move opposite each other. That is a statement about co-movement, and it is true — you can see it on the chart, and 2022 is the textbook case. This post is about forecasting, which is a different and much harder question, and the answer is that the DXY does not do it consistently at a 10-day horizon.
That is precisely why the macro backdrop shows a measured skill number next to every series instead of just drawing the line. A series can be worth watching for context while being worthless as a forecast, and the only way to tell the two apart is to score it.
What this does not prove
- Not "the dollar does not matter." It matters mechanically and it matters for context. It does not, on this label and horizon, reliably tell you the next ten days are unusually dangerous.
- Not "fade the dollar signal." Six years below 0.50 and a pooled 0.451 are tempting to invert. Our cross-fold standard deviation is 0.064 and a year is roughly 250 trading days, so individual years are noisy, and three years sit above 0.60 the other way. The claim we stand behind is the instability — a 0.235 spread across years around a coin-flip mean — not the direction of any one of them.
- Not a verdict on longer horizons. Ten days is short for a macro variable. A monthly or quarterly test is a genuinely different question and might well answer differently; we have not run it.
- A univariate screen. The DXY alone. Inputs that are weak on their own can still contribute inside a multivariate model.
The rule this series keeps arriving at
Three widely-watched indicators, three different ways of looking better than they are, one defence: ask for the per-period record, not the headline. A single number cannot distinguish "worked throughout" from "worked once", "tracks the market but never led it", or "works in some regimes and inverts in others" — and those demand completely different responses.
That is why w4rn publishes per-year skill next to every series, and why 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. Some of our numbers are modest. Showing them is the point — that is what calibration means.
Method note: figures from our univariate forward-AUC screen, combinatorial purged cross-validation, panel dated 13 September 2026, 2,281 observations for the DXY. Offline evidence, not a live trading claim.
Revision note, 14 September 2026: this post originally scored the DXY against SOL over 2020–2026 (0.494 pooled over 1,484 days; two good years, both the most recent). It now scores bitcoin over 2017–2026, the asset most readers mean by "crypto", with three more years of history. The longer record turns "the good years are the recent ones" into "the good years come and go" — a sharper version of the same warning.
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.