Research
Short-horizon mean reversion in cryptocurrency markets: a matched cross-market measurement
The authors ask a simple question with a demanding measurement design: at very short horizons, how much residual predictability is left in the two most structurally different markets available today…

- arXiv
- 2608.21888
- Published
- 2026-08-22
- Authors
- Nadav A. Kitron, Jonathan M. Wengrowicz
AI summary
Overview
Research area: Quantitative market microstructure and empirical asset pricing — specifically short-horizon return predictability (mean reversion), cross-market comparisons between cryptocurrency venues and US equity venues, and the informational content of exchange-traded wrapper prices.
Technical level: Advanced. The paper rests on a fixed, strictly out-of-sample walk-forward protocol, a constrained three-parameter logit, joint moving-block bootstrap inference, Benjamini–Hochberg FDR control, and an extensive artifact ledger. The findings themselves are stated in plain statistical terms (AUC, class-mean gaps, per-trade basis points), so a reader with intermediate statistics can follow the argument, but the methodological apparatus is specialist.
Scope (one sentence): The paper measures, under one matched out-of-sample protocol, whether 15-minute directional mean reversion is stronger and more pervasive in Binance crypto pairs than in liquid US stocks and ETFs, then uses NAV-linked listed wrappers and order-flow data to bound what that contrast means.
What This Paper Is About
The authors ask a simple question with a demanding measurement design: at very short horizons, how much residual predictability is left in the two most structurally different markets available today — 24/7 retail-heavy crypto spot versus exchange-listed US equities? They score 183 Binance USDT pairs and 187 US stocks and ETFs with identical models, identical walk-forward geometry, and identical inference, so the comparison is not a comparison of methods.
The paper then asks what the contrast is a property of. By studying US-listed funds whose net asset value is a crypto or metal price — and control stocks that are merely correlated with the same underlyings — the authors test whether short-horizon reversal attaches to the price process being wrapped or to the venue where the wrapper trades.
Key Contributions
-
A matched, population-scale cross-market measurement. The same primary test — universe definition, 15-minute bars from 2025-01-01 to 2026-02-11, train 5,760 / test 960 candle walk-forward, out-of-sample AUC, one-sided moving-block bootstrap (block 384, B = 300), per-asset conjunction of two models, and Benjamini–Hochberg FDR at q = 0.05 across all 370 assets — is applied to 183 Binance pairs and 187 US stocks/ETFs, with the design fixed before the wide universe was run.
-
A sign-level, model-free characterisation of the effect. The paper shows the crypto signal is a sign phenomenon rather than a magnitude phenomenon: lag-one return autocorrelation is near zero on major coins, yet a one-lag "bet against the previous candle" special case already reaches a crypto class-mean AUC of 0.527 against the twelve-lag kernel's 0.531. A parameter-free kernel-weighted sign correlation R (Eq. 2, with α ≡ 1 fixed a priori and nothing fitted) reproduces the class gap at −0.035 (95% CI [−0.041, −0.029]).
-
A wrapper-panel design that separates the price process from the listing venue. Eight underlyings (four OTC spot metals, Bitcoin and Ether on 24/7 exchanges, EURUSD and USDJPY on 24/5 spot FX) are matched to their NAV-linked US-listed funds — seventeen spot-custody wrappers plus the futures-based BITO and the 2× leveraged BITU and ETHU — alongside five correlated-but-not-NAV-linked controls (GDX, NEM, COIN, MSTR, UUP), all scored on identical New York 09:30–16:00 15-minute slots.
-
An explicit artifact ledger and a stated effect-size range. Sixteen mechanical or statistical channels that could manufacture the contrast are individually tested, plus a conjunction of the two most consequential, with the result reported as a range from +0.031 as fixed ex ante down to +0.011 with both artifact channels applied simultaneously.
Main Findings
-
The population contrast. Crypto AUC averages 0.531 with 98% of pairs above 0.5; stocks average 0.499 and split evenly around it. After joint FDR control, 90% of crypto pairs are significant against 2.7% of stocks (164 of 183 against 5 of 187). The joint moving-block bootstrap puts the class-mean gap at +0.031 (95% CI [+0.027, +0.035]), and an exact permutation null places the observed gap 4.4 null SDs out, beyond all 340 admissible label shifts.
-
The signal is small in size but broad and persistent. That is 3.1 points of AUC before artifact adjustment, 2.2 after, and 52.3% unthresholded accuracy on BTC. The authors are explicit that the effect is "not 183 independent replications": mean pairwise correlation of 0.40 gives roughly 2.5 effective independent crypto observations (stocks: 0.23 correlation, ≈ 4.2).
-
Reversal lives in signs, not magnitudes. The sign-flip rate rises monotonically across all ten deciles of the preceding move's absolute size, from 50.2% to 53.0%, and is above 50% in every decile. Because return autocorrelation weights bar pairs by the product of consecutive returns, the top decile's contribution nearly cancels the middle deciles' — which is why ρ₁ looks near zero. Stocks share the gradient at half the slope and from a lower base (49.8% to 51.2%, non-monotone).
-
Model-free corroboration. Wide-universe class-mean lag-one autocorrelations are −0.046 (returns) and −0.029 (signs) for crypto against −0.009 and −0.004 for stocks. The multi-lag sign autocorrelation is significant in 97% of crypto pairs versus 13% of stocks. R is negative in 98% of crypto pairs against 50% of stocks. Detrended fluctuation analysis gives crypto antipersistence with a class-mean Hurst gap of −0.049 (95% CI [−0.084, −0.012]).
-
Finite memory, standing effect. The signal decays to no-skill within hours (Figure 2); the hourly rerun keeps 58% of the 169 crypto pairs meeting the hourly bar-count cut significant after FDR versus 0% of stocks. Every focal coin-year since 2021 is two-model significant (24/24 coin-years), the sign correlation is negative in all 21 calendar quarters, and SPY scored identically never leaves the no-skill band.
-
The ex-ante universe is stronger, not weaker. Restricting to the 133 of 183 pairs already trading in the first sample month and re-ranking by that month's volume gives a class-mean gap of +0.035 (95% CI [+0.030, +0.038]) under the constrained logit and +0.030 [+0.026, +0.033] under the free logit; 97% of the 133 pairs are FDR-significant against the same 5 of 187 stocks. AUC is flat across ex-ante volume quintiles, with a slightly negative volume–AUC rank correlation (−0.16).
-
The frozen six-month holdout attenuates but survives. On 2026-02-12 to 2026-08-08 data — scored by a pipeline frozen against data through 2026-02-11, with only three sample-size thresholds relaxed — the gap is +0.020 (95% CI [+0.010, +0.028]) against +0.031 in-sample, and 60% of scored crypto pairs remain FDR-significant. The model-free R is nearly unchanged on the crypto side (−0.034 against −0.035 in-sample, negative in 96%), while the stock side moves from −0.000 to −0.006 (66% negative). Decomposing the attenuation: 23% is the universe change, 77% is decay on a fixed universe.
-
The equity null is powered, not merely empty. Median power to detect a crypto-sized effect across the 187 stocks is 0.98; power exceeds 0.8 in 94% of them, with only the shortest series falling to ≈ 0.64. The joint bootstrap puts a one-sided 95% upper bound of 0.5007 on the stock class-mean AUC — less than a fortieth of the crypto class mean's excess.
-
Both artifact channels together leave a gap clear of zero. Rescoring on non-flat bars only gives +0.022 (95% CI [+0.019, +0.025]); the one-bar feature/label gap leaves 71% of the excess; applying both simultaneously gives +0.011 (95% CI [+0.008, +0.014]), which is 36% of the headline gap — below the 47% independent channels would leave, indicating the channels overlap. Other checks: multi-venue refetches (Coinbase, OKX, Bybit) land 12/12 cells within ±0.002; quote-midprice bars leave BTC at 0.529 and ETH at 0.537; shuffled labels produce 0/370 significant; sign-randomized surrogates produce 0/14 significant.
-
Detectability is ruled out as the explanation. Intraday heteroskedasticity is 17.6× for listed instruments against 5.0× for crypto, and SPX — whose lag-one return autocorrelation of −0.030 is ten times BTC's — reads AUC 0.492 but lifts to 0.515 when its variance is homogenized. Recomputing AUC within hour-of-day buckets widens the gap from +0.0313 to +0.0335. BTC's ρ₁ of −0.003 supports only 0.501 through the linear channel against 0.533 observed.
-
One common factor does not explain it. Projecting
Authors’ abstract
At 15-minute horizons, directional mean reversion is far stronger and more pervasive in cryptocurrency markets than in US equities: scored under one matched, strictly out-of-sample protocol, 90% of 183 Binance pairs carry significant directional reversal against 2.7% of 187 US stocks and ETFs, in every focal coin-year since 2021. The signal lives in signs, not magnitudes: lag-one return autocorrelation is near zero on the major coins, yet simply betting against the previous…
The full abstract on arXiv