A 30-minute DQN does not finish, so it cannot beat the dummy

Verdict: NO_EDGE · Asset: BTCUSDT · Setup: Pro DQN on leftover WATCH features vs HIGH dummy rules vs price-only ablation · Budget: 200 episodes, 1800s cap, log reward, window 24, train_split 0.60, 5 bps next_open

Challenger, not fishing. Feature Lab already said leftover columns add nothing over a price MLP. This is the same question on the DQN surface: can the policy beat the HIGH dummy book?

What happened

Both trains cancelled at the 1800s wall. Leftover: episode 164/200, no report. Price-only: episode 165/200, no report. No model, no rl_evaluate, no costume check. Frozen run: we do not raise episodes after seeing the cancel.

Dummy 5 bps 8h short (recomputed, not copied): overlapping val n=450, +44.2 bp, t=4.10 (fact). Nonoverlap val n=71, t=1.42 (not a book). OOS overlapping t=1.69. Same object as lshighno / Feature Lab WATCH.

Q1 FAIL (null policy). Q1b missing. Q3 skipped.

Trading

Do not trade a cancelled DQN. Do not treat 164/200 as a near-miss to retune. The leftover HIGH dummy is still a level fact, still not a live book.

Product

Pro RL on ~52k 1h bars does not checkpoint when max_seconds fires. A paying user who trains for 30 minutes gets no model. That is a product hole, not a research leftover.

Final Verdict

NO_EDGE