Strategy harness v2

loading…

What 78 out-of-sample tests established

Each of these was a strategy selected for top-decile in-sample Sharpe, then measured on 2025–26 data it had never seen. They were spent by a bug — the adjudication loop had no ceiling and burned 78 looks against a budget of 10 — but the result is the clearest finding this programme has produced.

Intraday reversal — maker execution

The one real effect ever found here: a gross Sharpe of +1.2/+1.3, ruled out because a flat taker fee on 1,200–2,400×/yr turnover exceeded it. That is a real effect priced below cost, not an absent one. Reversal posts into the move — it is a natural maker, and MEXC maker is 0.00% against 0.02% taker. Pre-registered, judged by CPCV (28 purged out-of-sample paths) rather than the spent holdout, and reported across the whole cost-assumption grid. ideal is a deliberate upper bound with no adverse selection: if only it is positive, the result rests on a fill assumption kline data cannot settle.

Confirmatory results

The only track that can establish anything. A handful of pre-registered hypotheses — written down, with a stated prior reason and a falsification criterion, before being evaluated — each tested once against a bar calibrated for that handful. A hypothesis is admissible here only if it came from prior reasoning; promoting a strategy the search found scoring well would be selecting and confirming on the same data.

Exploratory search

Millions of specs. It has evaluated more than four million and confirmed nothing — and a positive control established that it never could: planting an edge of known size showed it needs a true Sharpe of ~2.51 for an 80% chance of detection. Turtle trading (+0.53) and a hypothetical Sharpe-1.2 strategy are both rejected. Its output is hypotheses and the placebo distribution, never findings.

AI review log

Every two hours the search is reviewed and the decision recorded here, at the time, permanently. The reviewer may add new strategy families; it may never drop one, reweight compute toward what is winning, or touch the gate, the bar or the holdout — selecting on past results was measured in this programme at −0.20, t = −3.1. Its job is to notice what the search space is missing, not to pick favourites inside it.

Where candidates die

v1 could only record pass or fail. v2 records how far each one got, so the funnel shows which test is doing the work.

Mechanisms

Every registered strategy family and what it has produced. The diagnosis behind adding these was that nine families shared roughly seven ideas — all of them "score a coin on its own price, rank the nine, long the top". A family here is pre-registered: a written prior reason, sha-stamped before its first evaluation. Deflated is the probability the true Sharpe beats what the best of this many searched attempts reaches by chance — the number that separates a finding from a lucky draw.

Closest to the bar

Ranked by margin over the noise bar in standard errors, not by Sharpe. v1's three published winners cleared by 0.00, −0.17 and −0.15 SE — which is why margin, not rank, is the column that matters.

Holdout ledger

The holdout is finite. Every look is recorded here permanently and raises the bar for the next one, so it degrades honestly instead of going stale in silence. v1 ran four million cases and spent zero looks.