Overview & Experiments 55 Synthesis Roadmap Lookahead Audit

Direction Forecast at 1h horizon — with CVD

Inconclusive

2026-05-20 directionlightgbmcvdshort-horizon

Hypothesis

Directional signal that vanishes at 4h (Mai-Welle direction_baseline IC +0.015) may appear at 1h horizon when CVD/order-flow features are added. LightGBM on the full Master-LGBM panel + CVD targets sign(fwd 1h log return). Promote if Spearman IC > 0.04 walk-forward + ≥70% stability + top-decile fwd return spread > 30 bps.

Verdict

**MARGINAL** — pooled IC +0.0232, marginally better than Mai-Welle's +0.015 (direction_baseline) but still below the production noise floor. 1h is better than 4h but still not enough.

n_windows

pooled_IC

+0.0232

ic_stability

+0.9231

top_3_features

['log_rv_1h_ann', 'ret_1h', 'ret_24h']

spread_stability

+0.9231

pooled_spread_bps

+2.0127

pooled_sign_hit_rate

+0.5084

Direction Forecast at 1h horizon — with CVD

2026-05-20 · status: inconclusive · 37.3s

Hypothesis: Directional signal that vanishes at 4h (Mai-Welle direction_baseline IC +0.015) may appear at 1h horizon when CVD/order-flow features are added. LightGBM on the full Master-LGBM panel + CVD targets sign(fwd 1h log return). Promote if Spearman IC > 0.04 walk-forward + ≥70% stability + top-decile fwd return spread > 30 bps.

Verdict: MARGINAL — pooled IC +0.0232, marginally better than Mai-Welle's +0.015 (direction_baseline) but still below the production noise floor. 1h is better than 4h but still not enough.

Key metrics

metric	value
pooled_IC	`+0.0232`
pooled_sign_hit_rate	`+0.5084`
pooled_spread_bps	`+2.0127`
n_windows	`13`
ic_stability	`+0.9231`
spread_stability	`+0.9231`
top_3_features	`['log_rv_1h_ann', 'ret_1h', 'ret_24h']`

Approach

Master-LGBM panel (now includes CVD features added 2026-05-20) with directional target = log(close_t+1h / close_t). LightGBM n_estimators=300, max_depth=5, learning_rate=0.05, min_child_samples=80. Walk-forward 12mo train / 3mo test, embargo 60 min. 13 valid splits.

Pooled OOS metrics

Spearman IC: +0.0232 (n=28,464)
Sign hit rate (predicted-sign matches actual): 50.84%
Top-Q5 vs Bot-Q1 conditional spread: +2.0 bps
Top-Q5 mean fwd return: +2.1 bps
Bot-Q1 mean fwd return: +0.1 bps

Per-window stability (13 windows)

IC > 0 in: 12/13 windows
spread > 0 in: 12/13 windows
IC mean: +0.0245, min: -0.0021, max: +0.0696

window	n	IC	spread_bps	sign_hit_rate
2023-01-01 → 2023-04-01	2160	0.0333	0.7	52.0%
2023-04-01 → 2023-07-01	2184	0.0405	6	50.6%
2023-07-01 → 2023-10-01	2208	0.0194	1.8	50.7%
2023-10-01 → 2024-01-01	2208	0.002	3.4	49.1%
2024-01-01 → 2024-04-01	2184	0.0263	-1.9	52.1%
2024-04-01 → 2024-07-01	2184	0.0696	6.9	52.9%
2024-07-01 → 2024-10-01	2208	0.0019	2.6	49.7%
2024-10-01 → 2025-01-01	2208	0.0167	0.2	50.9%
2025-01-01 → 2025-04-01	2160	0.0435	6.7	50.9%
2025-04-01 → 2025-07-01	2184	-0.0021	3.7	50.1%
2025-07-01 → 2025-10-01	2208	0.0212	1.8	50.6%
2025-10-01 → 2026-01-01	2208	0.0334	4	50.8%
2026-01-01 → 2026-04-01	2160	0.0129	1.4	50.6%

per-window

Top-15 features by mean LightGBM gain

feature	mean_gain
log_rv_1h_ann	199.2
ret_1h	177.8
ret_24h	172.7
ret_4h	165.2
log_vol	162.8
ret_7d	148.9
n_whale_trades_4h	146.7
range_4h	139.2
log_rv_1d_ann	137.8
bocpd_p_short	128.5
cvd_divergence_4h	127.4
vol_z_1d	124.2
vrp	119.5
hmm_p_state1	118.5
n_large_trades_4h	115.8

Overall verdict

MARGINAL — pooled IC +0.0232, marginally better than Mai-Welle's +0.015 (direction_baseline) but still below the production noise floor. 1h is better than 4h but still not enough.