AMD Power-of-3 Early Pull — walk-away

We set out to build a Strategy #8 around early-London AMD entries. The thesis was wrong, the inverse was real but 80% oracle-driven, and the live-safe residue was too thin to ship as a standalone. What we kept: Model 26 LON MOMENTUM BULL and a kill-switch.

The thesis we started with

Power-of-Three: Accumulation (Asia) → Manipulation (London sweeps Asia) → Distribution (NY reverses). If we enter in the AMD direction at the London sweep of Asia H/L, we’re front-running the reversal.

We coded it up (backtest_amd_po3_strategy.py), ran it on 6 years of NQ 1m data, and got this:

That’s not a weak result. That’s the thesis being catastrophically, reliably wrong. Every year. 13 standard deviations wrong.

Signal inversion, not a fake signal

A common trap: WRONG-DIR beats baseline, so you conclude the signal is noise. That’s the cargo-cult pattern. It’s wrong.

WRONG-DIR beating baseline by a huge margin in every OOS year means the real edge is the opposite of the thesis. PO3 as a framework is correct — the reversal does happen. It just doesn’t happen during the London session. It happens later, in NY. Enter at the London sweep and ride with it, not against it.

So we re-aimed at the WRONG-DIR side and ran a proper sanity battery (backtest_lon_momentum_deep.py).

The oracle trap

The deep battery showed +$39,675 and z=+12.25σ. Every DOW positive. Every year positive. Stop-tuning behaved (wider stops = better). This looked like a flagship edge.

Then the pattern broke one way: the whole result depended on a pure_only=True filter that selected only days where London swept just one side of Asia (high or low, not both). When both sides get swept it’s a chop day and longs and shorts both get stopped.

But you can’t know a day is “pure one-sided” until 05:00 ET closes. Our entry is at 02:xx. The filter is oracle.

Turning it off dropped the edge to +$7,054 over 5 years. Still positive — but +$1.4k/yr on 3 contracts doesn’t clear the bar we’ve set for building out a standalone NT8 strategy.

What live-safe filters look like

We ran entry-time-only features (observable at the sweep bar). Results:

The counterintuitive winner: entries at or below session VWAP for long sweeps. Mean-reversion. The more extended past VWAP at entry, the worse the trade — even though you’re entering with the momentum.

Why no NT8 build

Best live-safe composite:

sweep_hour == 02 AND time_to_sweep ≤ 15min
AND range_sofar < 20pt AND LONG ONLY
AND entry_close ≤ session_vwap

• n = 150 over 5 years (~30/yr)
• +$3,407 total, +$4,011 OOS in 3/3 years
• 53.3% WR

That’s ~$1,300/yr of OOS alpha per contract. The threshold we’ve set for spinning up a dedicated NT8 strategy is $2k+ OOS alpha and enough signals to distinguish edge from variance live. This clears neither.

What we shipped instead

1. Model 26 — LON MOMENTUM BULL — the composite above, ported as a Tier 2 PM line on our chart HUD. No strategy, just a probability target.
2. LON both-swept kill-switch — the 233 chop days (both Asia sides breached during London) had 7.3% WR and lost $32.6k. We now apply a −10 probability penalty to all London-window directional signals on those days.

Lessons

• Oracle filters are seductive. A peek-ahead filter turned +$7k into +$39k. Always verify a filter is live-computable before calling the edge real.
• WRONG-DIR > baseline = signal inversion, not a fake signal. The thesis is upside-down; the structure is real.
• NULL > baseline (no entry at all) is different: that means a filter is doing all the work and the thesis itself is cargo-cult.
• VWAP mean-reversion at the sweep was not something we looked for. We found it because we bucketed by entry-time features. Build your feature discovery before your thesis is locked in.

The walk-away is itself the artifact. We’d rather publish a dead-end writeup with the numbers than a shippable-looking strategy that dies OOS.