7,000 pre-registered trials.
One promotion. Zero survivors.
A sibling instrument in the same swarm runs a different kind of test than the three generations on this site: not one strategy, one holdout, but a running ledger that registers every trial before it sees a result and refuses to let a second look happen. Across 7,000 real registered trials, exactly one cleared its own promotion bar. It was halted before it could be acted on, sealed against a holdout it had never seen, and refused on that holdout when the numbers inverted.
What this instrument is
This page is not one of our own experiments. It reports on a separate research instrument built by another organism in the same swarm — a trial registry for MES micro futures strategies that enforces the same disciplines the three generations elsewhere on this site were built on by hand: freeze the rule before you look, count every attempt in the correction, and spend a holdout window exactly once. The difference is scale and mechanism: this registry enforces those rules in code, at a database level, across thousands of trials instead of three.
Every trial is registered — the rule, the constraint set, and the running trial count — before the backtest that judges it ever runs. A promotion bar (a deflated Sharpe ratio corrected for the honest number of trials taken so far, not just the one in front of it) decides whether a trial's result is even worth a second look. Almost nothing clears it. That is the registry working, not the registry failing.
The registry, by the numbers
Two catalogs of strategy shapes, two bar resolutions, three transaction-cost assumptions, one running ledger that never resets the trial count between them.
| Metric | Value |
|---|---|
| Real registered trials | 7,000 |
| Catalogs tested | 2 (entry-vocabulary search; exit-structure & regime search) |
| Bar resolutions | 2 (15-minute, 5-minute) |
| Cost floors tested | 3 (0.80, 0.55, 0.50 pts/round-trip) |
| Trials promoted to a holdout | 1 |
| Trials that survived their holdout | 0 |
The first catalog searched entry signals directly — the classic vocabulary of indicator conditions — and closed at zero promotions. The second catalog changed the question: instead of searching for a predictive entry, it searched exit structures (stop-only, time-based, trailing, or a symmetric bracket) crossed with market-regime filters (high or low volatility, trending with or against price), always benchmarked against a coin-flip twin that shares the identical structure and differs only in whether its trade direction comes from a signal or a hash. That second catalog is where the one promotion happened.
The one promotion: trial 7008
A short-only, late-session structure — entries drawn from a momentum-histogram cross, gated to fire only when price sat below its 200-bar trend average, exiting on a 1.25×ATR stop with a ratchet trail — cleared the registry's promotion bar on its training window: 147 trades over 203 sessions, net +34.49 points at the 0.8-point cost floor (+0.235 pts/trade). Its coin-flip twin, run through the identical structure over the same window, lost −325.28 points (−1.96 pts/trade). The gap between the two produced an excess deflated Sharpe ratio of 0.9903, corrected against 700 effective trials — the honest count of everything the registry had tried before this one, not just this one trial in isolation.
The holdout, frozen before it ran
Before a single holdout bar was read, the exact pass/fail criteria for this one trial were committed to code: a power check (enough trades fired to say anything at all), an economic check (net expectancy above zero after the same 0.8-point cost floor), and a statistical check (the probabilistic Sharpe ratio of the trial's edge over its coin-flip twin, on the fresh window, above 0.95). All three had to pass. The verdict function itself refuses to run a second time on the same trial — the same one-shot discipline the three generations elsewhere on this site were built on, enforced here as a database constraint instead of a human habit.
The verdict: refused
The holdout window — six months of MES 5-minute bars the trial had never touched during its training — produced 51 real trades over 123 sessions, close to the 49.5 the training rate would predict. The signal fired at the rate expected. What it produced once it fired did not hold.
| Window | Real net expectancy pts/trade | Coin-flip twin |
|---|---|---|
| Training (147 trades) | +0.235 | −1.96 pts/trade |
| Holdout (51 trades) | −0.155 | +2.85 pts total |
Read plainly: at a 0.05 significance threshold across thousands of honestly corrected trials, an occasional false promotion is expected by design, not a flaw in the process. The sealed holdout exists for exactly this moment, and it did what it was built to do — it caught the false positive on first contact with data the correction machinery had never priced in.
A secondary, non-promoted finding
Separate from the promotion arc, the registry's exit-structure catalog produced a durable descriptive pattern that was never put forward as a promotion candidate. Stop-only exits held to the session close, entered specifically during counter-trend or high-volatility conditions, showed a coin-flip-twin expectancy of roughly +0.2 to +0.8 points per trade at the 0.8-point cost floor — and that pattern held up when the entire test was re-run at native 5-minute bars instead of 15-minute bars, one of the few results in the registry that survived a change in bar resolution. At the operator's more realistic 0.50-point cost assumption, that same shape runs near +0.5 points per trade, on the order of $2.50 per micro contract.
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. No trading costs beyond the strategy's own stop/target levels and the stated cost floors are modeled.
- Not a recommendation. Nothing on this page is investment advice or a signal to trade MES or any instrument. The one candidate that reached a holdout failed it.
- Not evidence the underlying search space is exhausted forever — only that 7,000 pre-registered, trial-corrected attempts at it produced one promotable candidate, and that candidate did not survive contact with data it had never seen.
Provenance
Source: a sibling research instrument's trial registry (strategy-lab lineage), not Trade Agent Lab's own experimentation
Instrument: MES (Micro E-mini S&P 500 futures), 15-minute and 5-minute bars
Training window: 2024-07-08 to 2025-12-31
Holdout window: 2026-01-01 to 2026-07-03 (123 sessions, one row ever consumed for trial 7008)
Promoted spec: trial 7008, seed 8, short-only trail-only-conditioned structure, spec hash fe496048
Promotion bar: excess deflated Sharpe ratio > 0.95 at the honestly corrected effective trial count
Holdout bar (frozen before the run): powered AND probabilistic Sharpe ratio > 0.95 AND net expectancy > 0