Strategy Harness

→ Findings paper: the two strategies that survived 3.2 million tests

Every hypothesis the harness has tested, live. It runs 24/7 on a dedicated box, proposes strategies from a fixed grammar, and scores each one net of MEXC fees at both maker (0.00%) and taker (0.02%) rates. Nothing is selected by hand. The final bar is measured, not assumed: 12% of everything tested is synthetic noise, and a strategy must beat the Sharpe those random strategies are expected to reach at this sample size.

The funnel — how many survive each test

Winner board — everything that has ever passed

Sorted by Sharpe, but do not read it as a ranking. The columns that matter are on the right: how many years it worked, its worst year, and how its neighbours scored — the same idea at nearby parameter settings. A real effect has a whole neighbourhood working. A one-hit wonder is a lone spike surrounded by failures, and no amount of Sharpe fixes that.

Hover any column heading for what it means.

Previously passed, since demoted

The bar rises as more strategies are tested — testing more things means more chances for luck, so the evidence required goes up. These cleared it once and no longer do. Nothing was deleted and no number changed; only the threshold moved.

Nearest misses — reached the final bar and lost

Every one of these beat fees, worked in bull and bear, stayed market-neutral and cleared the trade minimum — then lost to the noise bar. Bar is the Sharpe a synthetic random strategy is expected to reach given how many strategies have been tested; it is measured from the injected placebos, not assumed. Gap is how far short the strategy fell.

Live log — most recent evaluations

By family