Strategy harness · findings · September 2026
An automated search tested 3.2 million trading strategies on nine crypto assets over nine years. Two survived. This is what they do, and why almost nothing else did.
A strategy had to clear every one of these before it counted. The bar was fixed in advance and applied identically to every candidate.
12% of everything tested is deliberately fake. Random signals, run through the identical pipeline. They exist to answer one question: how well does a strategy with no edge at all score on this data, given that we are testing millions of them?
The answer is sobering. The best random strategy reached a Sharpe ratio of 1.46, made money in both rising and falling markets, was profitable in 92% of six-month periods, held no market correlation, and stayed profitable at five times the real fee. Every conventional test of a trading strategy can be passed by pure chance if you try enough times.
So the final bar is not a judgement call: a real strategy must beat the score the fakes are expected to reach at this sample size. That threshold currently sits near +1.22, and it rises as more strategies are tested, because more attempts means more chances for luck.
Both are the same underlying idea at two different settings. Both are market-neutral: they hold long and short positions of equal size, so they make money from which coins do better than others, not from crypto rising.
xs_mom lb14 br4 rb21 /volatile
What it does. Every 21 days, measure how each of the nine coins performed over the previous 14 days — but first strip out whatever Bitcoin did. What remains is the coin's own performance, independent of the market. Buy the four strongest, short the four weakest, equal money on each side. Only do this when markets are volatile; sit out when they are calm. That works out to roughly 11 trades a year, in the market 45% of the time.
| Metric | Value | Meaning |
|---|---|---|
| Sharpe ratio | +1.86 | return per unit of risk |
| Annual return | 37% | compounded |
| Worst loss | −18% | peak to trough |
| Calmar | +2.01 | return per unit of pain |
| Rising markets | +2.37 | Sharpe when BTC above its 200d average |
| Falling markets | +0.98 | Sharpe when BTC below it |
| Correlation to BTC | +0.04 | effectively independent |
| Profitable years | 7 of 7 | 2018 through 2024 |
| Worst year | +0.22 | 2018 — never had a losing year |
Year by year: 2018 +0.22 · 2019 +2.92 · 2020 +2.80 · 2021 +1.71 · 2022 +1.39 · 2023 +1.07 · 2024 +2.47
xs_mom lb14 br4 rb21 /trending
What it does. Identical — 14-day market-adjusted performance, four long, four short, rebalanced every 21 days — but active when the market is trending rather than when it is volatile. About 10 trades a year, in the market 44% of the time.
| Metric | Value | Meaning |
|---|---|---|
| Sharpe ratio | +1.70 | lower than the first |
| Annual return | 33% | compounded |
| Worst loss | −18% | peak to trough |
| Rising markets | +1.86 | |
| Falling markets | +1.77 | far more balanced than the first |
| Six-month periods positive | 91% | vs 83% for the first |
| Correlation to BTC | +0.06 | effectively independent |
This is arguably the better of the two despite the lower headline. Its returns barely change between rising and falling markets (+1.86 against +1.77), where the first strategy leans heavily on rising ones (+2.37 against +0.98). It is also profitable in more six-month stretches. A lower Sharpe earned evenly is worth more than a higher one earned in one regime.
| Stage | Remaining |
|---|---|
| Strategies tested | 3,211,125 |
| Made money before fees | 1,831,552 |
| Still made money after fees | 1,630,705 |
| Worked in both rising and falling markets | ~290,000 |
| Consistent across six-month periods | ~92,000 |
| Market-neutral, enough trades, fee-robust | ~85,000 |
| Beat the noise bar | 2 |
The collapse at the last step is the finding. Eighty-five thousand strategies passed every test a human would normally apply — profitable, consistent, market-neutral, robust to fees — and were still indistinguishable from randomness once you account for how many were tried.
Nine liquid crypto assets, daily bars, August 2017 to the present. Every strategy is a combination of: what to rank the coins on, how many to hold on each side, how often to rebalance, and which market conditions to be active in. Signals use only price, volume and order-flow data available at the time of the decision; positions are lagged one day so nothing can act on information it would not have had. Fees are charged on every change in position at both the maker rate (0.00%) and the taker rate (0.02%), and results are shown at both. Results before 2025 are used for testing; 2025 onward is held back entirely.
The search runs continuously on a dedicated machine and has no memory of what has worked: it never steers toward promising families, because every attempt in this project to select on past results has made out-of-sample performance worse, not better.