We scored the Crypto Fear & Greed Index. Over eight years it's a coin flip.
Short answer: it doesn't forecast. We scored the Crypto Fear & Greed Index out-of-sample against a forward drawdown label on bitcoin across 3,142 days, from the index's launch in early 2018 to now. Pooled, it comes out at an AUC of 0.527 — where 0.50 = coin flip and 1.00 = perfect. Averaged year by year it is 0.485. Four of nine years sit above the flip, five below, and the one year that looks like a real signal — 0.640 — is 2025, the most recent complete year in the sample. Below is the method, the full year-by-year record, and — the part most write-ups skip — what this honestly does not prove.
What the index is
The Crypto Fear & Greed Index is a 0-100 score of market mood, from Extreme Fear to Extreme Greed, built from a blend of volatility, momentum, volume and survey inputs. It is probably the single most-screenshotted number in crypto: it gets quoted in newsletters, printed on dashboards, and used as a rough buy-the-fear heuristic by a lot of people.
It is not unstudied — there are academic papers on it, and plenty of backtests of the obvious strategy (buy at Extreme Fear, sell into Greed). But almost all of that asks would this have made money, which bundles the indicator together with an entry rule, an exit rule and a market regime. The narrower question is the one that decides whether the gauge is worth watching at all: does a Fear & Greed reading rank a risky day above a calm one, out of sample, and does it keep doing it? That number we could not find published anywhere, so we computed it — the same discipline we apply to our own readings: if we are going to show you a number, we owe you its track record.
How we scored it
We treat the index the way we treat every other input to our risk model — as a candidate predictor, screened on its own, with the result reported whether it flatters us or not.
- The question. For each day, does the index's level tell you anything about whether bitcoin is heading into a rough stretch over the next 10 trading days?
- The label. A 10-day forward drawdown in the worst 30% of historical 10-day windows. This is a research screen, not the number the cockpit shows you — our headline drawdown read is a much tighter event (a drop from today's price, at its worst point, within 1 day). Different question, different label; we are not quietly swapping one for the other.
- The base rate. 30% of windows qualify, by construction. So a useless predictor scores 0.50, and "it was right 70% of the time" would be a meaningless brag.
- The direction. Fear counts as the warning, matching how the index is read — a low reading says the crowd is scared. Above 0.50 means fear preceded the rough windows; below 0.50 means greed did.
- The scoring. Combinatorial purged cross-validation, with an embargo equal to the forecast horizon so no fold can peek across the boundary into its own future. Every number below is out-of-sample.
- The sample. 3,142 days, February 2018 to September 2026 — two full bear markets, two full bull markets, and the chop on either side.
If the phrase AUC is new to you, we wrote a plain-language explainer: what AUC means, and why we put it next to every forecast. The one-line version is that it measures ranking skill — given one risky day and one calm day, how often does the indicator rank them correctly?
The result
Pooled across the whole sample, the Crypto Fear & Greed Index scores AUC 0.527. Its precision-recall score is 0.305 against a 30% base rate — which is another way of saying it adds almost nothing to simply knowing how often bad windows happen. And it is a remarkably steady nothing: the cross-fold standard deviation is 0.022, the lowest of all 81 series we screen. Whatever the index is doing, it does it the same way in every slice of the data.
The year-by-year record is where it gets interesting (above 0.50 = fear preceded the rough windows; below = greed did):
- 2018 (from February) — 0.358
- 2019 — 0.453
- 2020 — 0.536
- 2021 — 0.541
- 2022 — 0.387
- 2023 — 0.477
- 2024 — 0.460
- 2025 — 0.640
- 2026 (to 13 September) — 0.517
Mean of the years: 0.485. Best: 0.640. Worst: 0.358. Notice which years sit where. The two years where the index ran hardest backwards — 2018 and 2022 — are the two bear markets, the years "buy the fear" actually paid: the rough windows more often followed the greedier readings than the fearful ones. The one year it looks like a genuine warning gauge is 2025.
In 2025 the index looked like a genuinely useful risk indicator. An AUC of 0.640 is better than almost anything in our own feature set. If you screened it at the end of that year, wrote it up, and built a rule around it, the backtest would have looked great and you would have felt clever.
It had scored above 0.55 in none of the seven years before.
The trap this illustrates
This is the cleanest real-world example we have of what we call the recency trap: an indicator whose apparent skill is concentrated in one regime — and the regime that flatters it is the one you would be discovering it from.
The mechanism is not mysterious. Fear & Greed is built substantially from volatility and momentum — it largely describes what price has just done. What that description means for the next ten days depends entirely on the market you are in. In a grinding bear market, a fearful reading marks a washout, and the next stretch is as likely to be a bounce as a further leg down; that is why 2018 and 2022 score well below the flip. In a market where the fearful stretches kept getting worse before they got better, the same reading is an honest warning; that is what 2025's number says. The index did not change. The market's response to fear did — and a gauge that moves with price cannot tell you which of those markets you are in.
This is the same distinction we keep coming back to: co-movement is not prediction. An indicator can track the market beautifully and forecast nothing at all. Pooling the whole history — which is what a single headline AUC does — averages the good regime and the bad ones into a number that hides the story; here it even lands on the flattering side of the flip, 0.527 against a yearly mean of 0.485, because cross-year pairs (a greedy 2021 day against a fearful 2022 day) get counted too. That is exactly why we publish per-year skill alongside the pooled figure rather than just the pooled figure.
What this does not prove
We would be doing the same thing we are criticising if we overstated this, so here are the limits, plainly:
- It is not "buy the fear, it's proven." Two bear-market years land well below 0.50, which is tempting to read as the contrarian rule confirmed. Resist it. You only know a year was a bear market after it happened, and the same reading scored 0.640 the other way in 2025. A sign discovered after the fact is not a strategy.
- Single years are noisy. Our cross-fold standard deviation is 0.022 on the pooled figure, but a single year is roughly 250 days, and a year moving from 0.46 to 0.52 is inside the range you would expect from chance. The claim we stand behind is the pooled one: a coin flip either way you average it — 0.527 pooled, 0.485 by year — with the one strong reading sitting in the latest complete year. Not "it has inverted", and not "it works now".
- This is one label on one asset. Bitcoin drawdown over 10 days. A different horizon or a different coin could score differently, and we have not tested every combination — our first version of this post scored SOL and found the same coin flip with a different best year, which is rather the point.
- This is a univariate screen. It measures the index alone. An input that is useless by itself can still contribute inside a multivariate model, and this screen is not the same thing as our full risk model.
- It is not useless as context. Knowing the crowd is fearful tells you something about positioning and liquidity. It just is not a forecast, and the gap between "interesting context" and "predicts risk" is the entire subject of this post.
Fear & Greed is not unusually bad — that's the real finding
It would be easy to read this as a hit piece on one index. It isn't, and the broader numbers are the reason.
We run this same screen over 81 series — rates, credit, cross-asset volatility, the dollar, funding, flows, sentiment. Across all of them the median AUC is 0.472. Only four clear 0.55. Fear & Greed ranks 12th of 81, which puts it comfortably in the upper half.
Read that again, because it is the actual story: a coin flip is an above-average result for a single indicator. Almost nothing predicts a forward drawdown on its own. The handful of series that do clear the bar do it by a hair, and none of them is a signal you would want to trade alone.
This is why we are suspicious of any tool built on one number, ours included. Useful risk estimation comes from combining many weak, noisy, partly-overlapping inputs and then calibrating the output — checking that when the model says 30%, the thing happens about 30% of the time. A single gauge, however well designed, cannot get there. Not because its authors did anything wrong, but because the market does not hand out that much predictability to one variable.
So the fair criticism of Fear & Greed is not "it's a bad index." It is that it gets used as though it were a forecast, and it has never been published with the one thing that would let you judge that: a score.
How we actually use it
We still track the index — it is one of dozens of series in the macro backdrop, and it earns its place as context. What it does not get is a promotion it has not earned. In w4rn every series carries its measured skill next to it, so a coin-flip indicator is visibly a coin-flip indicator rather than a number that looks authoritative because it is rendered in a nice gauge.
That is the difference we care about. Not "our index is better than their index" — the honest comparison is between a number that has been scored and a number that has not. Ours are cross-validated and Platt-calibrated, which means 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 a strange thing to aim for, we wrote a whole post on why a good 30% forecast usually doesn't come true.
The takeaway is not that sentiment is worthless. It is that an indicator without a track record is a vibe, and you should ask for the track record — including from us. Every forecast we publish carries its out-of-sample skill beside it, precisely so you can decide how much weight it deserves. Some of those numbers are modest. We show them anyway.
Method note: figures from our univariate forward-AUC screen, combinatorial purged cross-validation, panel dated 13 September 2026, 3,142 observations, 81 scored series. It is offline evidence, not a live trading claim.
Revision note, 14 September 2026: this post originally scored the index against SOL over 2020–2026 (0.505 pooled, 2,156 days, one strong year in 2021 that never returned). It now scores bitcoin over 2018–2026 — the asset most readers mean by "crypto", and two more years of history. The conclusion did not change; the strong year did, which is the finding.
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.