Is crypto trending or chopping this month? 10 months of 4-hour data say you can't tell in advance
Published by vyx
Crypto alternates between trend months and chop months, and nothing we could see in advance told them apart. On 4-hour bars across the top 100 Hyperliquid markets, momentum won 5 of 10 months and mean reversion the other 5 (the clear trend months by z 2.8–7.9). The best market-wide read, breadth, called all 5 months it was fitted on right and still made −0.2% of risk per setup against +5.3% for always trading momentum; on months it never saw, −1.0% against +0.9%, and −9.5% against −1.4%. Even a perfect hindsight month-picker would have added only 1.7 points.
Open the Trend-Month Momentum desk: the volatile crypto markets moving up on 4-hour bars right now, and how much of the field is trending up or down.
The question
Every trader has lived it: the breakout rule that printed money for a month stops working, and the fade that kept losing starts to pay. The usual explanation is the market's regime — some months trend, some chop — and the usual advice is to read the regime and switch styles. We wanted to know two things on real data: do crypto months really alternate between trending and choppy, and if they do, is there anything a trader can see in advance that says which month this one will be?
Two rules, one way of counting
We ran two opposite rules over every market and every 4-hour bar, both only on markets volatile enough to move (an average range of at least 0.8% of price):
- Momentum — trade in the direction of the last 20 bars’ move, strongest moves first.
- Mean reversion — bet against the stretch from the 50-bar average, most stretched first.
- Each signal is a setup, resolved by what price did next: +1R if it reached one risk unit in the setup's favour before one against it within 16 bars (64 hours), −1R if the reverse, 0 if neither. 1R is 1.5× the 14-bar average range — a normal stop. A 0.11% round trip (fees plus half the spread) is charged on every setup.
- We report the average result per setup as a share of what it risked: +5.3% means the average setup made back 5.3% of its risk after costs; a minus sign means it lost that much.
Average net result per setup, as % of the amount risked, for each calendar month from Sep 2025 to Jul 2026 with at least 100 setups of each rule (February is split where the two windows meet). Where one bar is long, the other is usually short: the months take turns.
The numbers
| Momentum | Mean reversion | |
|---|---|---|
| Oct 2025 | −4.9% | −0.2% |
| Nov 2025 | 12% | −14% |
| Dec 2025 | −9.6% | 7.4% |
| Jan 2026 | 18% | −35% |
| Feb 2026 (to the 16th) | −4.8% | −5.2% |
| Mar 2026 | −13% | 3.9% |
| Apr 2026 | −5.8% | 2.4% |
| May 2026 | 11% | −23% |
| Jun 2026 | 16% | −19% |
| Jul 2026 | −11% | 6.8% |
Yes, the months alternate
Of the 10 months, momentum won 5 and mean reversion 5, and the winner changed in 6 of the 9 handovers from one month to the next. The trend months were loud: Nov 2025 (z 2.8), Jan 2026 (z 7.9), May 2026 (z 5.5), Jun 2026 (z 5.4), plus Feb 2026 (to the 16th), a tie that went to momentum by 0.5 points. The chop months were quiet: mean reversion won them, but only Mar 2026 and Jul 2026 cleared the usual bar for chance (|z| ≥ 2). In the newest window, Sep 2026 went to mean reversion (z −1.2). So the regime is real — the question is whether it can be called before the month is over.
3 reads that might call the month
We tested the three market-wide reads most often suggested for regime detection, each computed only from bars that had already closed and smoothed over the last 24 bars (4 days), so a month is never judged by its own future:
- Breadth — the share of markets trading above their own 20-bar average.
- Dispersion — how far apart markets’ 20-bar returns are (their cross-sectional spread).
- Leader–laggard spread — the top fifth of markets’ 20-bar return minus the bottom fifth’s.
Each read got ONE threshold and ONE direction (above the threshold → momentum, or above → mean reversion), picked on Sep 2025 – Feb 2026 from 34 choices (17 thresholds × 2 directions): the one that called the most months right, ties broken by result. Then every 4-hour bar's setups were routed to the rule the read named. The thresholds were then frozen and run on Feb 2026 – Jul 2026 and Sep 2026 – Oct 2026, which the fit never saw. August 2026, between them, is held back for future checks and was never read. Rules for the fit and the verdict were written down before any result was looked at.
Each read's routed result minus simply trading momentum every month, in points of risk per setup — left of the line is worse than never switching. Breadth was the read picked on Sep 2025 – Feb 2026 (it called the most months right); dispersion and leader–laggard spread are shown for the record.
The numbers
| Routed by the read, minus always momentum (points of risk per setup) | |
|---|---|
| Breadth · Sep 2025 – Feb 2026 (fitted) | −5.5% |
| Breadth · Feb 2026 – Jul 2026 (unseen) | −1.9% |
| Breadth · Sep 2026 – Oct 2026 (unseen) | −8.1% |
| Dispersion · Sep 2025 – Feb 2026 (fitted) | 1.7% |
| Dispersion · Feb 2026 – Jul 2026 (unseen) | −8.6% |
| Dispersion · Sep 2026 – Oct 2026 (unseen) | 3.4% |
| Leader–laggard spread · Sep 2025 – Feb 2026 (fitted) | −2.6% |
| Leader–laggard spread · Feb 2026 – Jul 2026 (unseen) | −2.0% |
| Leader–laggard spread · Sep 2026 – Oct 2026 (unseen) | 1.0% |
Calling the month right is not the same as making the money
Breadth did what a regime detector is supposed to do: on the months it was fitted on it named the winning rule 5 times out of 5. Its routed desk still made −0.2% per setup, against +5.3% for momentum every month. On Feb 2026 – Jul 2026 it called 3 of 5 months right and made −1.0% against +0.9%; on Sep 2026 – Oct 2026, −9.5% against −1.4%. Dispersion looked like the better read in the fit — +7.0% per setup on Sep 2025 – Feb 2026 — then routed 4 of 5 unseen months right and made −7.8%. No read beat momentum on both windows it never saw.
Average net result per setup, % of the amount risked, pooled over the months each rule won (Sep 2025 – Feb 2026 to Feb 2026 – Jul 2026). Momentum loses a little when it is wrong; mean reversion loses a lot.
The numbers
| The 5 trend months | The 5 chop months | |
|---|---|---|
| Momentum | 13% | −9.1% |
| Mean reversion | −22% | 4.4% |
This is the mechanism. Switching to mean reversion in a chop month gains about 13.5 points per setup over staying with momentum. Switching in a trend month by mistake costs about 35.3 points — roughly 3× as much. A read has to be close to perfect before switching pays, and a read built on trailing data is late by construction: it flips after the move that defined the month, and on the sharpest bars of a trend month — exactly where mean reversion loses most — it is often still pointing the wrong way.
Sep 2025 – Feb 2026, net result per setup. The perfect month-picker knows each month's winner in hindsight and trades only that rule; "last month's winner" trades the rule that won the month before. The dashed line is always trading momentum.
The numbers
| Net result per setup, % of the amount risked (Sep 2025 – Feb 2026) | |
|---|---|
| Perfect month-picker (hindsight) | 7.0% |
| Always momentum | 5.3% |
| Routed by breadth | −0.2% |
| Routed by dispersion | 7.0% |
| Routed by leader–laggard spread | 2.8% |
| Last month’s winner | −22% |
| Always mean reversion | −14% |
Even with perfect hindsight, picking the month on Sep 2025 – Feb 2026 beat always-momentum by only 1.7 points per setup (+7.0% against +5.3%). On Feb 2026 – Jul 2026 the room was bigger — 6.5 points — and no read captured any of it. Following last month's winner was the worst choice of all (−21.7% on Sep 2025 – Feb 2026, −3.9% on Feb 2026 – Jul 2026): the months alternate rather than persist, so last month's style is usually the wrong one.
Every month, both rules
| Month | Momentum | Mean reversion | Winner | z | Setups (mom / rev) |
|---|---|---|---|---|---|
| Sep 2025 | −0.6% | −18.2% | too few setups | 1.4 | 90 / 154 |
| Oct 2025 | −4.9% | −0.2% | Mean reversion | −0.5 | 116 / 280 |
| Nov 2025 | +11.7% | −14.0% | Momentum | 2.8 | 139 / 264 |
| Dec 2025 | −9.6% | +7.4% | Mean reversion | −1.7 | 116 / 303 |
| Jan 2026 | +17.5% | −34.8% | Momentum | 7.9 | 296 / 545 |
| Feb 2026 (to the 16th) | −4.8% | −5.2% | Momentum | 0.0 | 103 / 188 |
| Feb 2026 (from the 16th) | −13.4% | +13.8% | too few setups | −1.5 | 67 / 54 |
| Mar 2026 | −12.6% | +3.9% | Mean reversion | −2.1 | 196 / 565 |
| Apr 2026 | −5.8% | +2.4% | Mean reversion | −1.1 | 247 / 467 |
| May 2026 | +11.4% | −23.1% | Momentum | 5.5 | 354 / 576 |
| Jun 2026 | +16.1% | −19.5% | Momentum | 5.4 | 323 / 532 |
| Jul 2026 | −11.0% | +6.8% | Mean reversion | −2.8 | 287 / 682 |
| Sep 2026 | −5.6% | +3.4% | Mean reversion | −1.2 | 241 / 433 |
| Oct 2026 | +20.0% | −28.2% | too few setups | 3.1 | 47 / 119 |
Net result per setup, % of the amount risked. z is the gap between the two rules over its standard error; |z| ≥ 2 is the usual bar for a gap chance rarely makes. A month counts when both rules have at least 100 setups.
The 3 reads, as fitted
| Read | Threshold (percentile of the fit window) | Above it, trade | Months right · Sep 2025 – Feb 2026 | Months right · Feb 2026 – Jul 2026 |
|---|---|---|---|---|
| Breadth | 0.419 (60th) | Mean reversion | 5 of 5 | 3 of 5 |
| Dispersion | 0.0572 (20th) | Momentum | 4 of 5 | 4 of 5 |
| Leader–laggard spread | 0.14 (15th) | Momentum | 4 of 5 | 2 of 5 |
change(close, 20) / close > 0 atr(14) / close >= 0.008
The first line is the direction the momentum rule traded — up over the last 20 four-hour bars (the study took both sides; the screen shows the long side). The second is the volatility gate every setup in the study passed. The study's own rule went further — it took only each market's strongest 20-bar moves against its own history — so the screen lights more markets than the rule traded.
These are setups resolved at ±1R after a modelled cost, not trades: no position sizing, no overlapping positions, no slippage beyond the modelled spread. Momentum itself was no edge here — +5.3%, +0.9% and −1.4% per setup on the three windows, none clearing the bar for chance. The finding is about switching, not about momentum: in this data, reading the month and switching styles did worse than not switching. Nothing here says to buy or sell anything.
How this was done
The same method as every study here — how vyx tests a signal.
- Venue and data
- Hyperliquid perpetuals, 4-hour candles: frozen snapshots of the venue's 4h bars for today's top 100 markets by volume over 13 months (the 4h history the venue serves is limited, so these windows cannot be re-downloaded — re-runs use the same snapshots, on request).
- Markets
- The top 100 Hyperliquid markets by volume when the snapshots were taken; 71 of them traded in Sep 2025 – Feb 2026, 90 in Feb 2026 – Jul 2026, 100 in Sep 2026 – Oct 2026.
- Windows
- Fitted: Sep 2025 – Feb 2026. Unseen: Feb 2026 – Jul 2026 (the thresholds were frozen first) and Sep 2026 – Oct 2026. August 2026 is a locked holdout, never read.
- Setups
- A setup is a 4-hour bar where a rule fired on a market, sampled every second bar. It wins +1R if price touches +1R before −1R within 16 bars, loses −1R on the reverse, 0 if neither; 1R = 1.5× the 14-bar average true range. Every setup pays a 0.11% round trip, converted to R through its own range.
- Month winner
- The rule with the higher average net result per setup that calendar month (UTC); a month counts when both rules have at least 100 setups.
- Reads
- 3 market-wide reads (breadth, dispersion, leader–laggard spread), each from closed bars only, averaged over the last 24 bars; one threshold and one direction per read chosen on the fitted window from 34 choices; a month is called by the rule most of its bars were routed to.
- Verdict rule
- Written before the run: the chosen read must call every fitted month right AND beat both rules on the fitted window, then beat both rules on each unseen window with at least 100 routed setups. It failed the first step; the unseen windows are reported for the record.
- Reproduce
- The study re-runs on request from the same frozen snapshots: every number on this page is regenerated by the same run, and the desk linked above shows the very momentum rule that was measured.
What it does not show
- Survivorship: the markets are today's top 100 by volume. A market that faded or was delisted during the 13 months is not in the data, and the rules may have behaved differently on it.
- One venue (Hyperliquid), one bar size (4 hours), 13 months — 10 months that counted in the fitted and first unseen windows. 10 months is a small sample for a monthly effect; a longer history could still hold a read that works.
- 3 reads were tried, each with 34 threshold-and-direction choices; reads built differently (per market rather than market-wide, on order flow, on volatility) were not part of this study.
- Setups, not trades: each setup is scored alone at ±1R with a modelled cost; real fills, overlapping positions, funding and sizing were not modelled.
- August 2026 was held back and never read, so the unseen windows are two, not three.
Questions
Is crypto trending or ranging right now?
Over the 13 months we measured on Hyperliquid, months alternated between the two: momentum won 5 and mean reversion 5 of 10. You only know which one a month was once it is over — none of the market-wide reads we tested called it early enough to make money from switching.
Can market breadth predict a crypto trend?
Not usefully, in this study. Breadth — the share of markets above their own 20-bar average — called all 5 fitted months right, yet routing by it made −0.2% of risk per setup against +5.3% for always trading momentum, and it lost to momentum again on both windows it never saw.
Should I switch between momentum and mean reversion with the market regime?
In this data, switching cost more than it earned. A right switch to mean reversion in a chop month gained about 13.5 points per setup; a wrong one in a trend month cost about 35.3. Unless the call is nearly perfect, staying with momentum did better. This is a study of setups, not advice.
Does crypto momentum work on the 4-hour chart?
It depends on the month. Momentum made +12.8% of risk per setup in the trend months and −9.1% in the chop months; over whole windows it came to +5.3%, +0.9% and −1.4% — not reliably above zero after costs.
What does R mean in trading?
R is the amount a trade risks — the distance from entry to the stop. A result of +1R made back what was risked; −1R lost it. Here every setup risks 1.5× its 14-bar average range, and results are averaged per setup and shown as a percentage of R.
Why publish a result that found nothing?
Because "read the regime and switch styles" is common advice, and testing it honestly is the only way to know whether it pays. A null result with a clear reason — the lopsided payoff — saves traders from fitting a regime filter that looks right in hindsight and loses going forward.
Watch it on today’s board
Open the Trend-Month Momentum desk: the volatile crypto markets moving up on 4-hour bars right now, and how much of the field is trending up or down.
Open the deskFree, no install. Or grade your own signal the same way — how vyx tests a signal.