A new indicator. 1,600 trials. Zero survivors.
After 49,263 trials had already run across the same sibling registry's existing vocabulary, spanning every instrument, session length, and resolution it had tried, the registry did something it had never done before: it added a new indicator. Not a new market. Not a new structure. One new piece of condition vocabulary, checked against a formal noise gate before it ever touched real data, then run through eight pre-registered strategy families and 1,600 real trials on the original instrument. Zero cleared the promotion bar.
A different axis, not a fourth leg
This page is not one of our own experiments. It reports on the same sibling research instrument behind our MES registry piece, gold leg, and crude oil leg. Those three held the strategy vocabulary fixed and varied the instrument. This one is different on purpose: it holds the instrument fixed (MES, the original one) and varies the vocabulary itself, by adding one new indicator that had never existed in this registry before. It is a sibling to that registry arc, not a fourth leg of it.
The new indicator is called gap: an overnight-gap state filter, a
per-bar condition comparing a session's own opening print against the immediately prior session's
closing print. It is the first indicator added to this registry's vocabulary since the catalog was
originally built.
Five candidates, one cheap enough to build safely
Before any new indicator code landed, a design study checked five candidate ideas against the actual backtesting engine, not against what sounded reasonable on paper. Four were set aside, each for a specific, code-checked reason:
- Order-flow-shaped bars (volume-delta proxies): would need a new bar-level statistic the engine does not compute anywhere today, plus its own methodology decision about which proxy formula counts as honest on data that has no true order-flow field.
- VWAP-relative session-structure entries: no volume-weighted running-average function exists anywhere in the backtest engine; building one is a real, separate piece of work.
- Multi-timeframe confirmation: the engine's own backtest function reads exactly one bar array; there is no mechanism anywhere in it to align a second, coarser timeframe against the first.
- Asymmetric or scaled exits: would need an entirely new exit-structure type plus new partial-fill accounting inside the fill engine itself, the largest build of the five by a clear margin.
The overnight-gap filter won by elimination: it fits the exact same "per-bar boolean held constant" shape two of the registry's oldest indicators already use, needed no new statistic, and needed no new data beyond a value the engine already computes for other purposes. Before it touched a single real trial, it still had to clear its own calibration gate: 200 trials against 250 sessions of pure noise, testing whether the promotion math itself produces false wins on data with no signal in it at all. Zero of 200 promoted on noise. That is what "cheapest to build" had to prove, not just claim.
A bug caught before it could matter
Part of this registry's honesty is publishing the near-misses, not just the results. While building the new indicator's own machinery, a real latent bug turned up: the piece of code that decides whether a trial gets the fuller structure and regime vocabulary checked for the literal text "v2", rather than checking whether the trial's version was v2 or newer. Left as written, every trial run under the new indicator would have silently fallen back to a thinner, older rule set that had never actually existed under this version's own label, one new indicator bolted onto an incomplete vocabulary instead of the full vocabulary the design study had actually frozen. It was found and fixed before a single real trial used it, and a dedicated test now checks for the exact failure mode directly.
The real spin: eight families, 1,600 trials, zero promotions
Eight pre-registered strategy families: four structure types (stop-and-hold, time-based exit, trailing-stop-only, and bracket order), each run twice, once with no regime filter and once with a regime condition applied, 200 seeds each, on native one-minute MES bars, against the same 0.8-point round-trip cost floor this registry's other MES families use.
| Metric | Value |
|---|---|
| Strategy families tested | 8 |
| Seeds per family | 200 |
| Real registered trials | 1,600 |
| Bar resolution | 1 minute (native) |
| Cost floor | 0.8 pts round-trip |
| Trials statistically promoted | 0 |
| Trials that reached a holdout | 0 |
| Family | Excess DSR |
|---|---|
| Stop-and-hold | 0.6582 |
| Time-based exit | 0.6730 |
| Trailing-stop-only | 0.4091 |
| Bracket order | 0.7842 |
| Stop-and-hold, regime-conditioned | 0.7351 |
| Time-based exit, regime-conditioned | 0.8398 |
| Trailing-stop-only, regime-conditioned | 0.7697 |
| Bracket order, regime-conditioned | 1.0 (see caveat below) |
All eight families closed clean: 200 of 200 seeds each, no trial in any of them reached the 0.95 promotion bar for real. There was no candidate to hold for a sealed holdout, because nothing in this program ever cleared the first bar.
One number that needs a caveat, not a headline
The bracket-order, regime-conditioned family's excess deflated Sharpe ratio of 1.0 looks, out of context, like the strongest number on this page. It is not a statistical signal. The specific trial that produced it took zero trades. Every candidate in this registry has to clear a minimum trade-count gate, 100 trades, before its Sharpe comparison is allowed to mean anything at all; a spec that never traded fails that gate outright, regardless of what the raw arithmetic on zero trades happens to produce. Two earlier families in this same program show the identical shape at smaller scale: the time-based-exit family's best nominal trial took 5 trades, and its regime-conditioned counterpart's best nominal trial took 1, both disqualified on trade count before the Sharpe number mattered. The one family in this program where the best trial actually cleared the 100-trade floor, trailing-stop-only with a regime filter, landed at an excess DSR of 0.7697 on 200 real trades, still short of the 0.95 bar. That is the cleanest like-for-like comparison this program produced, and it stayed a null too.
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 MES, gap conditions, or any instrument. No candidate in this program reached even a statistical promotion under trade-count-clearing conditions.
- Not evidence that new indicators can never help. Only that this one, specific, honestly-scoped addition to the vocabulary, tested across eight pre-registered families and 1,600 real trials on this instrument, produced no candidate that survived contact with the promotion bar.
Provenance
Source: a sibling research instrument's trial registry (strategy-lab lineage), not Trade Agent Lab's own experimentation
Indicator: gap (overnight-gap state filter), the first new indicator added to the vocabulary since the registry's original catalog build
Calibration gate: 0 of 200 promoted on a 250-session noise fabric, best DSR on noise 0.7367
Instrument: MES (Micro E-mini S&P 500 futures), 1-minute bars (native)
Training window: 2024-07-08 to 2025-12-31 (holdout window untouched by this program)
Real spin program: 8 structure families, 200 seeds each, 1,600 registered trials
Result: 0 of 1,600 real trials promoted; the one 1.0 excess-DSR reading belongs to a zero-trade trial that failed the 100-trade minimum