Overview & Experiments 61 Synthesis Roadmap Lookahead Audit

ETF-Flow Residual Check (20d) — Genuine Lead or Slow Momentum?

Inconclusive

2026-06-23 etfauditmomentumresidualoverlappositioning

Hypothesis

The scout lead etf_flow_5d → fwd_20d (IC +0.096, 3/3 years) is the only positioning claim not yet residualized. If, after orthogonalization against trailing price momentum (5d+20d) AND on non-overlapping 20d blocks, an IC >= 0.08 with CI_low > 0 remains, it is a genuine orthogonal slow lead; otherwise it is slow momentum / overlap-inflated.

Verdict

**PARTIAL / TOO THIN** — residualized, IC +0.126 (CI_low +0.045) pooled remains, BUT non-overlapping only +0.171 (CI [-0.231,+0.564], n=30). Effective sample too small for a deploy verdict; bear regime still missing.

yearly_resid_ic

{'2024': 0.137, '2025': 0.071, '2026': 0.251}

CI_residual_pooled

[+0.0452, +0.2121]

IC_flow_nonoverlap

+0.1382

IC_residual_pooled

+0.1263

IC_flow_5d_to_fwd20

+0.1002

ic_drop_pct_residual

-26.0303

IC_momentum5_baseline

-0.0187

CI_residual_nonoverlap

[-0.2306, +0.5643]

IC_momentum20_baseline

+0.0240

IC_residual_nonoverlap

+0.1706

effective_n_nonoverlap

r2_flow_explained_by_momentum

+0.4563

ETF-Flow Residual Check (20d) — Genuine Lead or Slow Momentum?

2026-06-23 · status: inconclusive · 2.2s

Hypothesis: The scout lead etf_flow_5d → fwd_20d (IC +0.096, 3/3 years) is the only positioning claim not yet residualized. If, after orthogonalization against trailing price momentum (5d+20d) AND on non-overlapping 20d blocks, an IC >= 0.08 with CI_low > 0 remains, it is a genuine orthogonal slow lead; otherwise it is slow momentum / overlap-inflated.

Verdict: PARTIAL / TOO THIN — residualized, IC +0.126 (CI_low +0.045) pooled remains, BUT non-overlapping only +0.171 (CI [-0.231,+0.564], n=30). Effective sample too small for a deploy verdict; bear regime still missing.

Key metrics

metric	value
IC_flow_5d_to_fwd20	`+0.1002`
IC_momentum5_baseline	`-0.0187`
IC_momentum20_baseline	`+0.0240`
IC_residual_pooled	`+0.1263`
CI_residual_pooled	`[+0.0452, +0.2121]`
IC_flow_nonoverlap	`+0.1382`
IC_residual_nonoverlap	`+0.1706`
CI_residual_nonoverlap	`[-0.2306, +0.5643]`
effective_n_nonoverlap	`30`
r2_flow_explained_by_momentum	`+0.4563`
ic_drop_pct_residual	`-26.0303`
yearly_resid_ic	`{2024: 0.137, 2025: 0.071, 2026: 0.251}`

1) Pooled ICs (overlapping 20d windows, n=585)

Relationship	IC	CI_low	CI_high	p
flow_5d → fwd_20 (scout replication)	0.1002	0.0188	0.1862	0.0153
flow_5d → trail_5 (confound: flow chases price)	0.4978	0.4275	0.5517	5.97e-38
trail_5 → fwd_20 (momentum baseline 5d)	-0.0187	-0.1019	0.072	0.652
trail_20 → fwd_20 (momentum baseline 20d)	0.024	-0.0633	0.1115	0.563
flow_resid → fwd_20 (orthogonal to momentum) ★	0.1263	0.0452	0.2121	0.00221

R² of flow_5d explained by trailing momentum: 45.6%
Residualization changes IC by -26.0% (+0.1002 → +0.1263)

2) Overlap-honest — non-overlapping 20d blocks

The overlapping sample n=585 is misleading: 20d windows overlap by ~95%. Disjoint blocks (every 20th day) → effectively n=30.

Signal	IC_nonoverlap	CI	p	n
flow_5d → fwd_20	0.1382	[-0.285, +0.521]	0.467	30
flow_resid → fwd_20	0.1706	[-0.231, +0.564]	0.367	30

3) Yearly IC (Spearman, in-sample per calendar year)

flow_5d → fwd_20: {2024: 0.04, 2025: 0.091, 2026: 0.126}
flow_resid → fwd_20: {2024: 0.137, 2025: 0.071, 2026: 0.251}

ic-comparison