1,100 gold trials.
One promotion. Zero survivors.
The same sibling registry that closed 7,000 MES trials at one promotion and zero survivors ran the identical vocabulary a second time, on a different instrument: micro gold futures. Eleven strategy families, 200 seeds each, native one-minute bars, 1,100 real registered trials. One cleared its own promotion bar. It never reached a holdout. A rule written before this candidate existed killed it first, on its own numbers, without spending the one-shot look this registry guards so carefully.
What this leg is
This page is not one of our own experiments. It reports on the same sibling research instrument behind our earlier MES registry piece, extended to a second live instrument: MGC, the micro gold futures contract. Same registry, same discipline, same promotion bar, a different market. If the durability of a null result matters, it should hold when the instrument changes and not just when the strategy vocabulary does.
Every trial in this leg was registered (the rule, the constraint set, and the running trial count) before the backtest that judges it ever ran, and benchmarked against a coin-flip twin sharing the identical structure, differing only in whether its trade direction comes from a signal or a hash. A promotion bar (a deflated Sharpe ratio, corrected for every trial the registry had already run, not just the one in front of it) decides whether a result is even worth a second look. Across this whole leg, almost nothing cleared it.
The leg, by the numbers
Eleven strategy families, mirrored from the same vocabulary already tested on other instruments in this registry, 200 seeds each, on native one-minute gold bars, against a fixed cost floor of 0.40 points round-trip.
| Metric | Value |
|---|---|
| Strategy families tested | 11 |
| Seeds per family | 200 |
| Real registered trials | 1,100 |
| Bar resolution | 1 minute (native) |
| Cost floor | 0.40 pts round-trip |
| Trials statistically promoted | 1 |
| Trials that reached a holdout | 0 |
| Trials that survived | 0 |
Ten of the eleven families closed clean: 200 of 200 seeds each, no trial in any of them came close to the promotion bar. The eleventh, a mean-reversion family, produced the leg's only promotion, on its fifth seed.
The one promotion: mean-reversion, seed 5
A mean-reversion structure cleared the registry's statistical promotion bar on its training window: 271 trades, an excess deflated Sharpe ratio of 0.9722 against its coin-flip twin at an effective trial count of 3,227 (the honest correction for everything the registry had tried before this one, not just this trial in isolation). That excess-Sharpe gap is exactly what the promotion bar tests for, and this candidate cleared it.
What the promotion bar does not test for is whether the candidate makes money. This one did not: its in-sample economics were a net −22.56 points over those 271 trades, roughly −0.083 points per trade at the stated cost floor. It looked statistically distinct from its coin-flip control while losing money in absolute terms.
Killed by a standing rule, not a judgment call
This registry keeps a one-shot holdout: a sealed window of price data a candidate is allowed to touch exactly once, spent only when spending it can change the answer. A rule already on the books before this candidate existed, first written for an earlier promotion on a different instrument's queue, says a statistically-promoted candidate whose own in-sample economics are already negative gets its holdout declined without being consumed. The reasoning is mechanical, not discretionary: a passing holdout would still describe a strategy that loses −0.083 points per trade. Spending the one look this registry has on that question has zero decision value.
So the holdout was never read for this candidate. The disposition on record is promoted-statistical, economically-negative, holdout declined, not consumed. The sealed window stays sealed, available for whichever future candidate actually needs it.
Ten families with nothing to kill
Outside the one promoted candidate, the other ten families in this leg closed at 200 of 200 seeds each, zero promotions, gate held clean every time. Best-trial deflated Sharpe ratios across those ten families ranged from roughly 0.09 to 0.89, all short of the 0.95 promotion bar; several of the nominal best trials were separately rejected on trade-count floors before the Sharpe comparison even applied. Nothing in this leg beyond the one mean-reversion candidate reached the promotion bar at all.
What this is not
- Not our own research. This page reports a sibling instrument's results as a source of record; it is not a strategy Trade Agent Lab built, ran, or endorses.
- Not a live result. Every number above comes from a historical simulation against recorded price data, not real order fills.
- Not a recommendation. Nothing on this page is investment advice or a signal to trade MGC or any instrument. The one candidate that was statistically promoted never even reached a holdout.
- Not evidence gold is unpredictable forever. Only that 1,100 pre-registered, trial-corrected attempts at this strategy vocabulary produced one statistically-promoted candidate, and that candidate's own in-sample economics were negative before a holdout ever entered the question.
Provenance
Source: a sibling research instrument's trial registry (strategy-lab lineage), not Trade Agent Lab's own experimentation
Instrument: MGC (Micro Gold futures), 1-minute bars (native)
Training window: 2024-07-15 to 2025-12-31 (holdout window untouched by this leg)
Promoted spec: mean-reversion family, seed 5, trial ID 32279
Promotion bar: excess deflated Sharpe ratio > 0.95 at the honestly corrected effective trial count
Disposition rule: pre-existing, applied verbatim (a statistically-promoted, economically-negative candidate has its holdout declined, not consumed)