Why I built this
I've been running a set of single-asset technical strategies. RSI, MACD, ROC, that kind of thing, tuned per coin. Earlier this week I finished a graduation-gate audit and retired 4,367 of 5,778 strategies because the specialist promotion lane had been letting through fluke backtests. What survives is a smaller, cleaner pool, but it is still cousins of each other in strategy space. All of it is single-asset technicals. If technical trading on individual coins stops working, everything I have stops working the same way at the same time.
Cross-sectional momentum is structurally different. Instead of asking "does BTC's RSI say buy?", it asks "over the last week, did BTC outperform SOL by more than SOL outperformed DOGE?" It ignores absolute price levels and indicator readings and just ranks assets against each other. Long the top, short the bottom. Hold. Rebalance.
The peer-reviewed and semi-peer-reviewed literature on this in crypto is fairly consistent. The Vilnius comparative study, the Setup Alpha 2026 edges guide, and a handful of other sources all cite a J=1 week / K=1 week variant as producing annualized Sharpe somewhere between 1.1 and 1.5. That's serious edge, and if it exists it would give me a bet that fails independently from what I already run.
What I built
The exact paper spec. No opinions layered on top.
- Universe: 10 majors that I already have full data for. BTC, ETH, SOL, XRP, DOGE, BNB, ADA, AVAX, LINK, LTC.
- Every 7 days: rank the 10 by the previous week's close-to-close return.
- Long the top 3, short the bottom 3, equal weight (1/6 of book each).
- Middle 4 sit out.
- Hold 7 days. Rebalance.
Data: daily closes from Binance spot, 2020-10-31 to 2026-09-02. 303 weekly rebalances.
Costs: 0.15% per leg (Binance perp taker plus spread), so 0.30% round-trip per position. Full turnover assumed (worst-case: every position rotates every week), which caps commission at roughly 1.8% at the book level per week. Real turnover is lower but I wanted the pessimistic case for the first read.
Funding cost omitted for v1. In a balanced long-short book, funding paid on longs and funding received on shorts largely offset, and this simplification would matter more if the aggregate result were borderline. It wasn't.
What came out
303 weekly rebalances. 5.8 years of data.
- Weekly return mean: 0.23%
- Weekly return standard deviation: 4.50%
- Weekly Sharpe: 0.052
- Annualized Sharpe: 0.37
- Probability the true Sharpe is greater than zero: 81% (need 95%+ to call this signal, not noise)
- Max drawdown: 41.7%
- Terminal equity from $1: 1.47 (roughly 6.8% per year net)
The annualized Sharpe I got is a quarter of the low end of the literature range and a third of the high end. That's not a small miss.
The per-year story is worse than the aggregate suggests.
2020: 8 weeks +60.5% wr 75% (tiny hot-start sample)
2021: 52 weeks +1.7% wr 42% (bull market — the edge disappeared)
2022: 53 weeks +19.4% wr 45% (bear — the edge showed up)
2023: 52 weeks +23.6% wr 50% (recovery — the edge kept working)
2024: 52 weeks +0.4% wr 44% (flat — the edge stopped)
2025: 52 weeks -24.0% wr 46% (broken)
2026: 34 weeks -11.1% wr 32% (still broken)
The last 86 weeks have been consistently negative. That's not a slow drift or a rough patch. It's a sustained regime in which the strategy loses money.
What I think this means
Two clean years out of six is not a durable edge. And even the two working years were not spectacular: 19% and 24% net, on a strategy running with 42% max drawdown and 45% win rate.
More importantly, if the edge is real and just quiet, you would expect the win rate to hover around 50% in dead periods and swing up in good periods. Instead the win rate is under 50% in 5 of the 7 years, and the 2026 win rate is 32%. That's the shape of a strategy that has been at least partially arbitraged out, not one that is dormant.
I don't know for sure why the edge decayed. Some plausible mechanisms:
- 2024-2025 saw crypto ETF flows, meaningful institutional participation, and a shift from retail-driven price action to something more like traditional index behavior. Retail chasing was the fuel for short-horizon momentum. If it thinned, the edge thins.
- The specific weekly lookback / weekly hold that the paper cites has been public knowledge for years. Any documented edge on public data eventually attracts capital that trades against it.
- My universe is too narrow. Cross-sectional edges want variance in the ranking; 10 large-caps that increasingly co-move don't give you much variance to rank against.
I could test each of those. But that walks straight into the trap that just ate 3,925 of my strategies earlier this week: keep sweeping variants until one looks good in backtest, then be surprised when it fails in forward. So I'm not going to.
What I'm doing instead
Two things.
First, I'm going to keep publishing these when they happen. This kind of result is more valuable to me than a paper describing an edge that supposedly exists, because it tells me what happens when you actually run the spec. If the whole industry is publishing "here is a strategy that worked in a curated window" and no one is publishing "I ran that strategy honestly and it lost," the incentive gradient distorts what people believe about the field.
Second, I'm pivoting the current research effort to delta-neutral. Basis trades and funding rate capture do not require me to be right about which way crypto is going next week. That's a smaller-payoff, higher-consistency edge, and it survives my strongest concern about the momentum result: even if crypto price action has become efficient, funding-rate arbitrage and cash-and-carry spreads are structural. They exist because leveraged perp traders pay for their leverage, and that fee has to go somewhere.
I'll write up those results the same way: what I built, what came out, whether it worked.
What I'd tell anyone building a similar thing
Build the paper spec first, exactly as described. Do not add your own opinions until you have the clean baseline result. If the baseline works, then justify each parameter change on paper before making it. If the baseline doesn't work, that's the answer.
The temptation to sweep is enormous, and the tell that you've overfit is that the "working" variant has a hyperparameter tuning story you invented after seeing the data. The paper you cited didn't need one; if yours does, you're modeling noise.
Backtest detail is saved at forward-test-data/xsmom-backtest-v1.json if anyone wants to reproduce or challenge the numbers.