Overview & Experiments 55 Synthesis Roadmap Lookahead Audit

Direction Forecast — 15min Horizon

Dropped

2026-05-20 directionlightgbm15minshort-horizon

Hypothesis

Going from 1h to 15min horizon either firms up directional IC (order-flow less absorbed) or dies in noise. Pooled IC > 0.04 + Q5-Q1 spread > 15 bps walk-forward = production-tradable above 11 bps fee floor.

Verdict

**NO IMPROVEMENT** — pooled IC +0.0229, spread +0.79 bps, hit rate 51.02%. Going shorter than 1h does NOT help — noise dominates. Direction prediction with free data is closed; need paid microstructure / LOB.

n_windows

pooled_IC

+0.0229

ic_stability

+0.9231

top_3_features

['log_vol', 'ret_15m', 'log_rv_1h_ann']

headroom_decile_bps

-10.0210

pooled_sign_hit_rate

+0.5102

headroom_quintile_bps

-10.2058

pooled_Q5_Q1_spread_bps

+0.7942

pooled_D10_D1_spread_bps

+0.9790

Direction Forecast — 15min Horizon

2026-05-20 · status: dropped · 6.4s

Hypothesis: Going from 1h to 15min horizon either firms up directional IC (order-flow less absorbed) or dies in noise. Pooled IC > 0.04 + Q5-Q1 spread > 15 bps walk-forward = production-tradable above 11 bps fee floor.

Verdict: NO IMPROVEMENT — pooled IC +0.0229, spread +0.79 bps, hit rate 51.02%. Going shorter than 1h does NOT help — noise dominates. Direction prediction with free data is closed; need paid microstructure / LOB.

Key metrics

metric	value
pooled_IC	`+0.0229`
pooled_sign_hit_rate	`+0.5102`
pooled_Q5_Q1_spread_bps	`+0.7942`
pooled_D10_D1_spread_bps	`+0.9790`
headroom_quintile_bps	`-10.2058`
headroom_decile_bps	`-10.0210`
n_windows	`13`
ic_stability	`+0.9231`
top_3_features	`['log_vol', 'ret_15m', 'log_rv_1h_ann']`

Approach

Panel at 15min cadence with 28 features. Target = log fwd 15min return. LightGBM per walk-forward fold. 13 valid folds.

Fee floor: Hyperliquid taker 0.045% × 2 (round-trip) + slippage ~2 bps = ~11 bps. Need Q5-Q1 spread > 15 bps + IC > 0.04 to be tradable.

Pooled OOS metrics

Spearman IC: +0.0229 (n=113,856)
Sign hit rate: 51.02% (coin flip = 50%)
Q5-Q1 spread (quintile): +0.79 bps
Q5 top: +0.56 bps | Q1 bot: -0.23 bps
D10-D1 spread (decile, more aggressive): +0.98 bps
D10 top: +0.68 bps | D1 bot: -0.30 bps

Per-window stability (13 windows)

IC > 0 in: 12/13 windows
spread > 0 in: 10/13 windows
IC range: [-0.0028, +0.0493], mean +0.0250

window	n	IC	spread_bps	sign_hit_rate
2023-01-01 → 2023-04-01	8640	0.0493	1.54	52.3%
2023-04-01 → 2023-07-01	8736	0.0357	1.72	51.0%
2023-07-01 → 2023-10-01	8832	0.0259	1.31	51.2%
2023-10-01 → 2024-01-01	8832	0.037	2.1	50.7%
2024-01-01 → 2024-04-01	8736	0.017	1.19	50.6%
2024-04-01 → 2024-07-01	8736	0.0388	1.8	50.8%
2024-07-01 → 2024-10-01	8832	0.0216	1.6	50.8%
2024-10-01 → 2025-01-01	8832	0.0197	0.25	50.9%
2025-01-01 → 2025-04-01	8640	-0.0028	-1.38	50.1%
2025-04-01 → 2025-07-01	8736	0.004	-0.38	50.5%
2025-07-01 → 2025-10-01	8832	0.0341	0.91	51.6%
2025-10-01 → 2026-01-01	8832	0.0106	-0.79	50.9%
2026-01-01 → 2026-04-01	8640	0.0339	0.95	51.7%

per-window

Top-15 features by mean LightGBM gain

feature	mean_gain
log_vol	223.2
ret_15m	211.8
log_rv_1h_ann	207.2
signed_volume	182.8
ret_1h	180.1
ret_4h	162.3
ret_24h	138
range_4h	133.2
n_whale_trades_4h	124.4
log_rv_1d_ann	123.2
log_rv_4h_ann	122.5
ret_7d	119.6
bocpd_p_short	117.5
vrp	103
n_large_trades_4h	101.5

Verdict

NO IMPROVEMENT — pooled IC +0.0229, spread +0.79 bps, hit rate 51.02%. Going shorter than 1h does NOT help — noise dominates. Direction prediction with free data is closed; need paid microstructure / LOB.

Fee math

Hyperliquid taker fee: 0.045% per side → 9 bps round-trip
Slippage estimate: ~2 bps (small BTC orders)
Total fee floor: ~11 bps round-trip
Quintile-trade headroom: -10.21 bps (negative = lose money on average even with perfect quintile selection)
Decile-trade headroom: -10.02 bps