Two rules, one name

MRdata · methodology note · 2026-07-26

On 2026-07-03 this site registered a pattern rule and started scoring it in public: resource specialists and investment banks both net-buying at least $5k the same session, volume at least 2× average, no price-sensitive announcement that day. Every qualifying hit would land on the tracker and have its 5-session forward return logged, win or lose.

That part worked. Between 07-03 and 07-25 the registered rule produced 6 qualifying hits — roughly 8 per month — and logged them: BOA, KTA, EXR, MSB, LRV, BRN. Four have resolved so far. The record is intact and stays published.

The drift

The failure is what happened around it. By late July, the page a reader actually saw — the pre-catalyst radar — was quietly running a different rule than the one being measured. Three separate divergences, none of them announced, each introduced for locally sensible reasons:

1. The quiet window. The registered rule requires quiet on the hit day. The displayed surface demanded no price-sensitive announcement across a much longer window.
2. The desk test. The registered rule requires the two specific desks — resource specialists and investment banks — both buying. An adjacent conditioner table accepted any two of three desk groups, including market makers, under the same visual framing.
3. The universe. The baseline these numbers were compared against included ~190 off-watchlist large-caps that ride along on the broker tape — the contamination corrected and restated on the track record on 2026-07-25.

The compounded result: the displayed cell qualified about 2.7 names per month. Two names surfaced in three weeks, and each dropped off the page within days when news landed. A reader watching the radar was watching a rule that was operationally dead — while the measured rule logged hits the reader mostly never saw. Displayed and measured had the same name and different definitions.

That is the second methodology failure caught here in one month, after the baseline universe contamination (published 34.6% win restated to 31.5% once the stray large-caps were removed). Both were found by internal audit, not by luck, and both are published in full rather than quietly patched — but they should not have happened, and the honest description is the one above.

The fix

Structural, not cosmetic:

• One implementation. The quiet gate now has a single implementation with the window as an explicit parameter. Two definitions can no longer circulate under one name.
• Displayed = measured, permanently. The radar now renders exactly the hits of the live registered rule. Nothing is filtered for display. RSI and drilling status are badges on hits, never gates. The rule's wording ships inside the exported data, so the page cannot describe a rule other than the one it renders.
• Three sequential periods, one timeline. Everything before 07-03 is calibration history — in-sample, explored, not evidence (42 hits, 42.9% win, median 0.0%). The 07-03 rule's receipt (6 hits, ~8/month) is genuine out-of-sample evidence under a stated, dated rule: it retires at the 08-04 re-registration with its record intact and visible. The revised rule's receipt starts empty on 2026-08-04. The old "frozen discovery sample" framing is deleted; its outcome data stays in the lake.

The recalibrated rule

From 2026-08-04: resource specialists and investment banks both net-buying at least $5k the same session · volume at least 1.5× average · no price-sensitive announcement in the 3 calendar days ending at the hit day.

Two changes from the 07-03 rule, and the reasoning is construct validity, not performance:

• Quiet window, 1 day → 3 days. Measured over every candidate name-day in the calibration window, the same-day quiet leg passed 91.8% of days. A leg that passes 91.8% of the time is decoration — and a rule whose quiet leg does not gate cannot honestly be called "positioned before the news." The 3-day window passes 79.7%: a leg that actually selects. Would this argument survive with the returns column hidden? For D, yes — 91.8% pass-through is a fact about the gate's selectivity, derivable from the leg-attribution table with outcomes covered up.
• Volume floor, 2.0× → 1.5×. In July's thinner tape the 2× floor collapsed the rule's frequency from ~26 to ~8 hits per month. At 1.5× the July-regime rate is roughly 16 per month — enough to accumulate a resolved out-of-sample sample in weeks rather than quarters. A rule that cannot generate a sample cannot be tested at all.

What the historical win rates of the candidate rules were played no part in this choice, and they are deliberately not reproduced in this note — recalibrating for measurability inside a declared calibration window is defensible; tuning until the backtest looked better would not be.

One subset is pre-registered alongside the rule: RSI under 40 at the hit. It is the only one. Any further slice would be a new registration, dated, or it does not count — a growing list of slices is multiple comparisons in slow motion. Wherever the subset is reported, its coverage is explicit: of N hits, M had RSI data, K of those were under 40. Indeterminate rows are counted in the denominator, never silently dropped.

The dates

• 2026-07-03 — first rule registered; its receipt runs to 08-04 and stays visible.
• 2026-07-25 — universe contamination corrected and restated (first failure).
• 2026-07-26 — this note: the drift documented, surfaces unified, revised rule declared (second failure).
• 2026-08-04 — revised rule re-registers with the v2 freeze; its receipt starts at zero.

Supersession, not deletion. Every number this note replaces remains published where it originally stood.

MRdata is not a licensed financial adviser. Nothing here is advice — it is a measurement log.