PREVIEW — real data, not yet part of the daily build. Unlinked page; forward-OOS window only (from 2026-05-20). Research classifications, not recommendations.

Signal Decomposition — forward-OOS panel

build e4087a0d · method b3c57f89911a57bb · protocol 826e82d83591 · 46 snapshot dates × 79 names · window 2026-05-20 → 2026-07-21 · primary endpoint: per-component IC at k=5 · benchmark-relative (vs SPY) forward returns · point-in-time scores

Per-component table

CompIC (95% CI)BadgePartialMarginal ΔICMonoChurnPlacebo pctile
C1+0.038 (-0.116,+0.181)DESCRIPTIVE-0.025+0.0220+0.700.9390.982
C2DATA-LIMITED (constant 0 across all live snapshots — under diagnosis)
C3-0.030 (-0.129,+0.099)DESCRIPTIVE-0.064-0.0179-0.200.4340.048
C4DATA-LIMITED (frozen/constant input — flagged, not scored)
C5+0.066 (-0.061,+0.196)DESCRIPTIVE+0.044+0.0127-0.300.9441.0
C6-0.063 (-0.121,-0.003)PRELIMINARY-0.032-0.0157-0.900.8910.0
C7+0.118 (-0.026,+0.253)DESCRIPTIVE+0.093+0.0338+0.900.861.0
C8+0.007 (+0.000,+0.021)DESCRIPTIVE-0.004+0.0000-0.400.00.994
C9+0.013 (-0.035,+0.059)DESCRIPTIVE+0.006-0.0046+0.100.9720.766

Placebo percentiles are two-sided extremity readings under a within-date permutation null — exploratory, not a significance claim.

Redundancy (|rho| > 0.6)

Pairrho
none above threshold

Horizon decay (IC at k=1/2/5/10/20)

Compk=1k=2k=5k=10k=20
C1+0.019+0.006+0.038-0.001-0.015
C3-0.021-0.065-0.030+0.006-0.042
C5+0.018+0.018+0.066+0.084+0.075
C6-0.075-0.095-0.063-0.037-0.031
C7+0.058+0.073+0.118+0.110+0.091
C8+0.004+0.005+0.007+0.002+0.001
C9+0.012+0.008+0.013-0.003-0.030

Per-class IC (k=5)

CompBROAD_ETFEQUITYMACROSECTOR_ETF
C1+0.038 (d=40)+0.076 (d=40)-0.024 (d=40)
C3-0.030 (d=40)+0.000 (d=40)-0.053 (d=40)
C5+0.067 (d=40)+0.000 (d=40)+0.000 (d=40)
C6-0.070 (d=40)+0.000 (d=40)+0.000 (d=40)
C7+0.141 (d=40)+0.000 (d=40)+0.000 (d=40)
C8+0.008 (d=40)+0.000 (d=40)+0.000 (d=40)
C9-0.028 (d=40)+0.048 (d=40)+0.120 (d=40)

Cells of exactly +0.000 within MACRO / SECTOR_ETF are constant-cross-section artifacts — the component does not differentiate names within that class, so no rank correlation exists; they are not measurements of zero signal. The BROAD_ETF column shows "—" because the class (5 names) is below the module's 8-name cross-section floor: no statistic was computed, which is not the same as zero.

Research and education only. Not investment advice. Component diagnostics are research classifications, not recommendations. No forward alpha is established.