Close near the high fades a little, then fees eat it

Verdict: NOT_TRADEABLE · Asset/TF: BTCUSDT 1h (primary), 15m (secondary) · Sample: chronological 60/20/20

Hypothesis: close location inside the bar loc = (close−low)/(high−low) predicts the next bar — either continuation (near high → next up) or fade (near high → next down).

What happened

Frozen thresholds: loc ≥ 0.8 / ≤ 0.2; events only when range > discovery median.

• Continuation (loc≥0.8 → next up): rejected on validation (1h and 15m).
• Fade (loc≥0.8 → next down; loc≤0.2 → next up): clears validation with the same frozen cuts.
• 1h val (range-filtered): P(up|high) ≈ 0.449 vs mid 0.526 vs unc 0.509; P(up|low) ≈ 0.580. Same fade shape on 15m.
• Q2: loc is ~97% collinear with sign(r_t), but within sign strata loc still adds a fade increment — so not a pure “just the last bar’s sign” story, and not a clean NO_EDGE.
• OOS: weaker on 1h high side; 15m more stable.

Trading

Q3: fade edge present at 0 bps; dies at 5 bps round-trip. Not tradeable under realistic costs.

Final Verdict

NOT_TRADEABLE

There is a small mean-reversion in where the bar finishes, beyond raw sign, but it is too thin for 5 bps. Do not ship a “close at extreme → fade” bot on BTC klines without a venue where costs are far below that.