Research Note 05

Reversal: Two Kinds of Mean Reversion, One Closed on Sight

The note frames its research question around a naming collision: the literature contains two genuinely different strategies both called "reversal," neither arriving with a ready-made, retail-executable recipe, and the note's job is to keep them separate, close one at the literature stage, and leave the other open for replication.

Branch A — Dobrynskaya

Branch A follows Dobrynskaya (2021/2023, Journal of Alternative Investments), tested on roughly 2,000 coins above a $1 million market-cap floor on CoinMarketCap-aggregated daily data from 2014 to 2020. Sorting by a one-to-two-week trailing return and holding the loser basket for ten to twelve weeks produces a reversal effect the paper itself frames as bubble-and-burst behavior, with a long-only practical version favored since shorting a broad tail of small coins is difficult. Two problems in the paper's own methodology sharpen the closure: there is no delisting or survivorship treatment beyond the entry-point market-cap floor, meaning the loser leg — the one leg this strategy would concentrate a portfolio into — is the least checked for what happens when a position goes to zero; and the 2014–2020 sample predates the crypto market's largest well-known permanent wipeouts (Terra/Luna, FTX, Celsius), so the underlying claim was never tested against the exact scenario a skeptical trader should worry about most.

Branch A is closed without an independent replication, on the same discipline applied to Grid and Trend: the paper's own diversified, dozens-to-hundreds-of-coins academic construction has no honest, unmodified translation into a retail-sized, concentrated portfolio, and shrinking it down would test an invented strategy wearing the paper's citation, not the paper's actual finding.

Branch B — Bianchi, Babiak & Dickerson

Branch B follows Bianchi, Babiak & Dickerson (2022, Journal of Banking & Finance), tested on cryptocurrency pairs against USD/USDT from March 2017 to March 2022 across five exchanges including Binance. Coins are split daily by de-trended trading volume, and within the low-volume group a short-term, market-making-style reversal is applied: buy what fell yesterday, sell what rose, hold one day, rebalance daily. The paper's own exchange-by-exchange breakdown (its Table 7) is unusually specific: on Binance, net-of-fees returns flip negative (−0.213% per day equal-weighted, −0.341% value-weighted, both significant), while on the other four exchanges the net-of-fees result stays positive and significant, reaching +0.811% per day net on GateIO equal-weighted.

Branch B is left open rather than closed, because it is concrete and well-specified enough to actually replicate on current data: a single-exchange, one-day holding-period strategy with no multi-month conviction call on any coin's survival, and a paper that already tells the project where the effect held up net of fees and where it did not. A retail replication is planned on Gate.io, using current OHLCV, volume, and fee schedule rather than the paper's 2017–2022 sample, run through the same bootstrap, sign-flip, and permutation tests used for Momentum.

Full note: GitHub ↗