Null result — published as tested

We fixed the confound.
The result barely moved.

Gen-1's fixed-point entry looked like it was degrading as MES's price level drifted over two years. Gen-2 re-expressed the same entry idea in ATR-relative terms — removing that specific confound — and re-ran the one-shot holdout. The point estimate turned positive. It still didn't clear its own bar.

The pre-registration

The entry signal was redefined as a ratio to volatility rather than a fixed point count: W1 fires when |r_first| / ATR14 ≥ T1_atr, W2 fires on sign agreement with |r_day| / ATR14 ≥ T2_atr. Thresholds (T1_atr = 0.722, T2_atr = 2.971) were calibrated outcome-blind as medians over 365 training sessions (2024-07-03 to 2025-12-31) — a window that does not overlap the holdout. The stop/trail exit law, the one-at-a-time position rule, the fill model, and the coin-flip twin all carried over from Gen-1 unchanged; only the entry gate's form changed. Frozen in commit 6b41c4363204 before the holdout ran.

The result

123 holdout sessions, 2026-01-01 to 2026-07-03 — the six months immediately following the training window, never touched during calibration. 129 trades fired.

Gen-2 holdout, 123 sessions, ATR-relative entry
Leg	Trades	Hit rate	Total pts	Expectancy pts/trade
Real signal	129	51.2%	+38.25	+0.297
Coin-flip twin	129	48.1%	−83.75	−0.649

Paired difference (real − twin)
Pairs	Mean diff pts	SD	SE	t-statistic
129	+0.946	19.54	1.72	0.55

Pre-declared bar: t ≥ 2.0 and expectancy > 0.8 pts/trade. Actual: t = 0.55, expectancy +0.297. Positive on both legs of the comparison, and still a clear fail on both prongs of the bar — not close enough to call a near-miss. At the honest 0.55–0.8 point transaction-cost floor this research treats as the bar for a real-money edge, a live version of this rule nets roughly −0.3 to −0.5 points per trade after costs, not the +0.297 the raw backtest shows.

The window split cleanly in the same direction both halves: first 82 sessions +0.314 pts/trade, final 47 sessions +0.266 pts/trade. Unlike Gen-1's split result, this one is at least consistent — just consistently short of the bar.

What actually moved, and what it points to

Re-expressing the entry gate in ATR-relative terms did what it was supposed to do: the point estimate moved from clearly negative (Gen-1: −0.22 pts/trade) to weakly positive (Gen-2: +0.297). That is a real, if small, improvement from removing a real confound. It is not, on its own, evidence of an edge.

The detail that mattered next: the coin-flip twin lost −0.649 points per trade over this holdout, running through the exact same stop-and-trail exit as the real signal. A signal-free random direction should not reliably lose money against a fair bracket unless the bracket itself has an unfavorable expectancy on its own — which raised the question of whether the trailing-stop exit was the thing quietly costing money, independent of whatever the entry gate was doing. Testing the exit law in isolation, holding the entry gate fixed, became the next generation's one adaptation. See the Gen-3 result.

Combined record across all three generations tested so far: Gen-1 (fixed-point entry) t=−0.61. Gen-2 (this page, ATR-relative entry) t=0.55. Gen-3 (exit law only) t=−0.01, the sharpest null. None cleared its own pre-declared bar. Full arc at the methodology page.

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, one-shot holdout 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: 6b41c4363204ae78c6827e55060929cf609c3c60
Instrument: MES (Micro E-mini S&P 500 futures), 5-minute bars
Training window: 2024-07-03 to 2025-12-31 (365 sessions, outcome-blind calibration)
Holdout window: 2026-01-01 to 2026-07-03 (123 sessions, 129 fired)
Coin-flip twin: direction fixed in advance by hash of session date, run through identical stop/trail logic
Success bar (declared before the result): paired t ≥ 2.0, expectancy > 0.8 pts/trade

No affiliate relationship: we don't have an affiliate program with Autoview and don't receive a commission if you sign up through the link below.

Read the Gen-3 result Get Autoview free →