Overview & Experiments 55 Synthesis Roadmap Lookahead Audit

Divergence scan v2 — bigger pivots + rolling volume

Dropped

2026-05-25 divergencersivolumepattern_scanv2robustness

Hypothesis

v1's no-edge result may stem from (a) too-small k=5 pivots producing noisy swings and (b) single-bar volume being a poor proxy at the pivot. Hold everything else constant and sweep pivot_k ∈ {10, 20} and vol_window ∈ {1, 10}. If neither knob lifts any pattern above the PURSUE bar, divergences are confirmed dead on BTC.

Verdict

**DROP** — even with bigger pivots/rolling vol the best cell is only +127.3 bps (`reg_bull_rsi` k=10, vol_win=1, 4h/24h) with p=0.412. v1's null result holds across the full robustness grid. Divergences confirmed dead on BTC as standalone entry filter.

best_p_value

+0.4122

best_pattern

reg_bull_rsi @ k=10, vol_win=1, 4h/24h

best_signed_bps

127.25

best_total_signals

cells_with_p_under_10pct

best_windows_under_5pct_p

cells_with_signed_gt_10bps

Divergence scan v2 — bigger pivots + rolling volume

2026-05-25 · status: dropped · 7.8s

Hypothesis: v1's no-edge result may stem from (a) too-small k=5 pivots producing noisy swings and (b) single-bar volume being a poor proxy at the pivot. Hold everything else constant and sweep pivot_k ∈ {10, 20} and vol_window ∈ {1, 10}. If neither knob lifts any pattern above the PURSUE bar, divergences are confirmed dead on BTC.

Verdict: DROP — even with bigger pivots/rolling vol the best cell is only +127.3 bps (reg_bull_rsi k=10, vol_win=1, 4h/24h) with p=0.412. v1's null result holds across the full robustness grid. Divergences confirmed dead on BTC as standalone entry filter.

Key metrics

metric	value
best_pattern	`reg_bull_rsi @ k=10, vol_win=1, 4h/24h`
best_signed_bps	`127.25`
best_p_value	`+0.4122`
best_windows_under_5pct_p	`3`
best_total_signals	`39`
cells_with_p_under_10pct	`0`
cells_with_signed_gt_10bps	`42`

Approach

Grid sweep over the two methodologically weak parameters in v1:

pivot_k ∈ [10, 20]
vol_window ∈ [1, 10] (rolling causal mean ending at the pivot bar)

Everything else identical to v1: 14-period RSI, 50-bar max-lookback between pivots, walk-forward 12mo train / 3mo test / 24h embargo, 200-perm mask-permutation test per window. Pattern definitions and sign conventions also unchanged.

Per-cell scan

k=10, vol_window=1, tf=1h (55,826 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=243, reg_bull_rsi=198, hid_bear_rsi=166, hid_bull_rsi=191, reg_bear_vol=331, reg_bull_vol=289, hid_bear_vol=290, hid_bull_vol=294

k=10, vol_window=1, tf=4h (13,958 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=64, reg_bull_rsi=47, hid_bear_rsi=43, hid_bull_rsi=46, reg_bear_vol=74, reg_bull_vol=70, hid_bear_vol=72, hid_bull_vol=91

k=10, vol_window=10, tf=1h (55,826 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=243, reg_bull_rsi=198, hid_bear_rsi=166, hid_bull_rsi=191, reg_bear_vol=365, reg_bull_vol=289, hid_bear_vol=318, hid_bull_vol=309

k=10, vol_window=10, tf=4h (13,958 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=64, reg_bull_rsi=47, hid_bear_rsi=43, hid_bull_rsi=46, reg_bear_vol=94, reg_bull_vol=82, hid_bear_vol=73, hid_bull_vol=88

k=20, vol_window=1, tf=1h (55,826 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=92, reg_bull_rsi=57, hid_bear_rsi=44, hid_bull_rsi=52, reg_bear_vol=104, reg_bull_vol=85, hid_bear_vol=87, hid_bull_vol=84

k=20, vol_window=1, tf=4h (13,958 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=17, reg_bull_rsi=12, hid_bear_rsi=11, hid_bull_rsi=12, reg_bear_vol=22, reg_bull_vol=13, hid_bear_vol=20, hid_bull_vol=16

k=20, vol_window=10, tf=1h (55,826 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=92, reg_bull_rsi=57, hid_bear_rsi=44, hid_bull_rsi=52, reg_bear_vol=110, reg_bull_vol=81, hid_bear_vol=87, hid_bull_vol=79

k=20, vol_window=10, tf=4h (13,958 bars, 21 WF windows)

Signals (full sample): reg_bear_rsi=17, reg_bull_rsi=12, hid_bear_rsi=11, hid_bull_rsi=12, reg_bear_vol=28, reg_bull_vol=19, hid_bear_vol=17, hid_bull_vol=16

Pooled OOS results (sorted by signed mean)

pivot_k	vol_window	timeframe	horizon	pattern	direction	n_windows	total_signals	mean_signed_bps	median_signed_bps	mean_p_value	p_under_5pct
10	1	4h	24h	reg_bull_rsi	long	20	39	127.25	79.03	0.412	3
10	10	4h	24h	reg_bull_rsi	long	20	39	127.25	79.03	0.412	3
20	10	4h	24h	hid_bear_rsi	short	8	10	92.85	111.75	0.405	1
20	1	4h	24h	hid_bear_rsi	short	8	10	92.85	111.75	0.405	1
20	1	4h	24h	reg_bear_vol	short	13	17	76.98	16	0.468	1
10	10	4h	24h	hid_bear_rsi	short	18	41	76.93	13.53	0.418	0
10	1	4h	24h	hid_bear_rsi	short	18	41	76.93	13.53	0.418	0
20	1	4h	24h	hid_bull_vol	long	10	13	72.92	-26.98	0.487	1
20	1	1h	24h	hid_bull_rsi	long	19	39	66.72	103.66	0.418	3
20	10	1h	24h	hid_bull_rsi	long	19	39	66.72	103.66	0.418	3
20	10	4h	24h	reg_bear_vol	short	12	22	57.24	17.4	0.504	0
10	1	4h	24h	reg_bear_rsi	short	18	46	54.57	14.71	0.429	2
10	10	4h	24h	reg_bear_rsi	short	18	46	54.57	14.71	0.429	2
10	10	4h	24h	reg_bull_vol	long	21	71	51.15	31.22	0.428	1
20	10	1h	24h	hid_bull_vol	long	21	63	47.33	32.8	0.495	2
20	1	4h	4h	reg_bull_vol	long	10	13	45.64	19.26	0.344	0
20	1	4h	4h	hid_bear_rsi	short	8	10	40.66	6.55	0.425	1
20	10	4h	4h	hid_bear_rsi	short	8	10	40.66	6.55	0.425	1
20	1	4h	24h	hid_bear_vol	short	12	16	39.59	-9.67	0.46	1
10	1	4h	24h	reg_bull_vol	long	20	61	36.86	22.66	0.455	2

Full grid in results_full.csv.

Best signed-bps per (k, vol_window) cell

pivot_k	1	10
10	127.25	127.25
20	92.85	92.85

Top 15 across grid