OPEN EXPERIMENT · PUBLISHED BEFORE THE WINDOW

Can a varied ML team beat the original pair?

Registered 11 September 2026 at 19:23 UTC. Standalone publication: 12 September 2026 UTC. Runs 14 September 00:00 UTC to 19 September 00:00 UTC.

A frozen diverse ML panel conditioned on confirmed Micro/Intraday turns improves net one-hour payoff over the retained intraday pair and matched controls.

Frozen models, a shared schedule, and matched controls. No changes to fit the result. This experiment runs separately from the live account and does not test its trailing stop.

All six criteria must pass.

1. At least 40 resolved active diverse-panel observations on at least four UTC dates
2. At least 90% of scheduled open-session grid decisions recorded on time
3. Positive base AND stress net pips for diverse panel
4. Positive mean paired net-pip advantage over intraday pair, no-level and no-compressor controls at both cost sets
5. Diverse panel beats both always-long and always-short at both cost sets
6. One-sided 95% day-block bootstrap lower bound for diverse-panel minus intraday-pair mean is positive; 2000 resamples, seed 20260914

How the comparison works

Same non-overlapping hourly decision grid and future entry for every panel. Ablations retrain the six new models with the named input groups removed; setup timing stays fixed and is not ablated. Intraday pair is a frozen retained-artifact control. No adaptive refitting in this experiment.

First M5 opening strictly after scheduled decision; decision grid is UTC hour at minute 05, using only bars completed by that time. Record before entry. No hindsight catch-up observations.

Isolated fixed clip(3*recorded ATR,5,50) pips; adverse gap/stop-first OHLC accounting; one-hour exit. Does not test live SMA trailing or account portfolio. · Simulation; account sizing N/A.

Costs and interpretation

Base: 1.5-pip spread, 0.2-pip slippage per side, 0.5 pip/day financing. Stress: 3, 0.5 and 1 respectively. Simulation; account sizing N/A. A separate spread sweep keeps slippage and financing fixed; a joint-cost result is not a spread threshold.

Any unmet criterion is not_passed, including insufficient activity, missing coverage, fitting/freeze failure or missing observations. Publish all controls and criteria, not only winners. No post-hoc extension.

Small prospective screening experiment, not proof of durable edge, calibrated win probability, GA benefit or live profitability.

Publication commitment

The existing independent experiment monitor publishes its result after close plus 70 minutes. The outcome page gives passing and failed results equal prominence. If the monitor is unavailable, that absence is disclosed; no success is inferred.

Download the public registration summary. Private training seeds and artifacts are not distributed.

Original registration SHA-256: 5300f15b061f977f863003c90a9172486a4de151a5564170f5769714ac81e447