NQ · model #14

AMD

Accumulation → Manipulation → Distribution. Asia establishes range, London sweeps one side, NY trades toward the opposite. Tier 1 touch magnet — 65% within 120 min.

updated 2026-04-23
tier 1 both liquidity-magnet amdpo3londonliquidity-magnetict-core

The setup

Asia session (20:00–00:00 ET) quietly builds a range. London (02:00–05:00 ET) sweeps one side of that range — typically the side opposite the eventual NY bias. New York then trades toward the other extreme of the London move.

If London sweeps Asia high → bearish manipulation → NY targets London low. If London sweeps Asia low → bullish manipulation → NY targets London high.

The line drawn on our heads-up display (HUD) is the touch target — it marks where NY is expected to visit, not necessarily close past.

Entry rules

The signal gets a time-decay haircut as the window burns:

Minutes inTouch-rate multiplier
0–301.00× (≈65%)
30–600.40× (≈26%)
60–900.22× (≈14%)
90–1200.10× (≈6%)

If it’s going to run, it runs fast. Median touch time is 3–4 minutes.

Backtest

n = 1,013 ONLY_H / ONLY_L sessions (5-year MNQ 1-min):

WindowTouch rate% of eventual hits captured
30 min52.6%72%
60 min59.1%81%
120 min64.9%89%
180 min67.9%93%
390 min (full RTH)73.1%100%

The re-audit story

An earlier audit looked at AMD as a close-direction predictor — and it failed (47–53%, pure noise). That almost got the whole signal demoted.

But “high probability” was the wrong frame. AMD isn’t a close-direction signal; it’s a touch magnet. Price visits the target at some point during the session, then does whatever it wants. Different question, different answer.

The old 88–92% figure was from a flawed baseline. The 65% at 120 min number above is the correct one, measured against proper conditional-remaining-touch rates.

Full writeup: Touch vs close direction — two questions, two signals (coming soon)

Kill-switch interaction

If both Asia high and low get swept during 02:00–05:00 ET, the day is a chop day (lon_both_swept=True). A −10 probability penalty is applied; lines that drop below prob 40 are removed. AMD is one of the penalized signals.

Regime-conflict penalty (2026-04-23)

A follow-up audit (n=1,105 AMD signals in the backtest CSV) asked the opposite question: does AMD break down when the broader regime points the other way? Base rates are strong — BULL 94.5% (n=582) / BEAR 90.2% (n=523) by the touch-rate metric in the results file — but two conflict axes chew into that:

Gap state (RTH open vs prior day H/L):

SignalGap stateHit ratenΔ vs base
AMD BEARABOVE_PDH81.7%115−8%
AMD BULLBELOW_PDL81.6%87−13%

Gap × 4H trend (the combined cell, not additive):

SignalCellHit ratenAction
AMD BEARABOVE_PDH + 4H BULL77.4%93−13 prob
AMD BULLBELOW_PDL + 4H BEAR76.6%64kill-switch

The implementation lives in PredictionModel.cs / PredictionModelWeb.cs immediately after the lon_both_swept block, and in backtest_prediction_model.py at the Model 14 consume site. Penalties are non-stacking: the matrix cell is a direct measurement, not a sum of independent effects.

Friday DOW also showed degradation in this audit (BULL 72.7% / BEAR 68.4% vs 91–100% Mon–Thu). The −20 Friday penalty that was removed on 2026-04-22 as a “flawed metric” has been reinstated in both PredictionModel.cs / PredictionModelWeb.cs and remains in backtest_prediction_model.py — the touch-rate measurement shows the DOW effect is real.

Full audit: C:\SMC\backtest_results\amd_regime_audit.txt.

Why Tier 1

Even at 65% touch with aggressive time decay, this is one of the highest-confidence lines on the board. The touch edge is stable across DOW, years, and regime. The “magnet” nature means it’s an honest touch signal even when the market has no directional conviction.

History