19 sessions looked real.
488 said no.
The first generation of this research lineage tested a fixed-point momentum entry on MES micro futures. A small calibration-window replay was encouraging. A full out-of-sample replay against two years of data was not. This is the story of why the second number is the one that counts.
The pre-registration
The entry law was a fixed-point breakout: enter when price moves a fixed number of points from a
reference level (T1 = 22.25 points, T2 = 32.75 points), with a stop sized at 1× the 14-period ATR
(floor 6 points) and a trailing stop activated after 1× ATR of favorable movement. Thresholds were
calibrated outcome-blind from 19 real RTH sessions of MES tick data (2026-03-22 to 2026-04-17) and frozen
in commit d3990f638bb3 before any holdout test ran.
What the 19-session window showed — and why it wasn't counted
Before the real out-of-sample test, the frozen strategy was replayed through the same 19-session archive used to calibrate it — partially circular, and explicitly flagged as such at the time. The number looked good: 52.6% hit rate, +2.24 points per trade on the real signal against −5.62 for a coin-flip twin, a paired edge of +7.86 points per trade.
The out-of-sample result
488 real trading sessions, 2024-07-03 to 2026-07-03 — two years of MES 5-minute bars, the calibration window excluded. 237 trades fired.
| Leg | Trades | Hit rate | Expectancy pts/trade |
|---|---|---|---|
| Real signal | 237 | 48.1% | −0.22 |
| Coin-flip twin | 237 | — | +0.63 |
| Mean diff pts | SE | t-statistic |
|---|---|---|
| −0.85 | 1.40 | −0.61 |
Why the 19-session number was a lucky pocket
The two results aren't measuring different things gone wrong — the same code, run on the calibration-window slice of the 488-session archive, reproduced the 19-session replay almost exactly (+7.86 points per trade paired, t=1.2, totals 42.25 vs. 42.5 points across two independent data sources). The pipeline is not the problem. The 19-session window was a real, if small, pocket of favorable conditions, and the honest two-year sample shows it did not generalize.
One likely contributor, flagged but deliberately not patched mid-generation: T1/T2 are fixed-point thresholds over a two-year window in which MES's price level moved roughly from 5,000 to 7,500. A fixed number of points is a shrinking percentage move as the price level rises — a probable source of the degradation. Re-testing that as an ATR-relative threshold, properly pre-registered rather than patched in place, became the next generation's one adaptation.
What this is not
- 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 are modeled.
- Not a recommendation. This page describes what one specific, frozen rule set did in one specific out-of-sample window. It is not investment advice and not a signal to trade MES or any instrument.
- Not evidence the underlying idea is wrong forever — only that this specific, pre-registered attempt at it did not clear its own bar, on this data, in this window.
Provenance
Freeze commit: d3990f638bb3dca514e72d35dd5443c2b5dd4b4e
Instrument: MES (Micro E-mini S&P 500 futures), 5-minute bars
Calibration window: 19 RTH sessions, 2026-03-22 to 2026-04-17
Out-of-sample window: 2024-07-03 to 2026-07-03 (488 sessions, calibration window excluded)
Coin-flip twin: direction fixed in advance by hash of session date, run through identical stop/trail logic