Result — 112 large caps, 447 sampled days, 2016–2026
| score correlation | IC correlation | |
|---|---|---|
| signals in catalog | 31 | 31 |
| effective breadth | 8.40 | 5.57 |
| √breadth (the multiplier) | 2.90× | 2.36× |
| √31 if independent | 5.57× | 5.57× |
| top-3 eigenvalue share | 53.9% | 65.5% |
| mean |correlation| | 0.207 | 0.300 |
| clusters at |r| ≥ 0.5 | 6 (one holds 24) | 5 (one holds 25) |
Twenty-five of the thirty-one signals form a single cluster. Momentum,
reversal, moving averages, RSI, Bollinger, drawdown depth, close location,
volatility and skew are all chained together at |r| ≥ 0.5. Only
negative_skew and rev_1d stand alone, with vol_trend/volume_shock and
illiquidity/small_dollar_vol as two-name pairs — and that second pair is
one of the duplicates planted as a control.
Some correlations are near-identities the labels concealed: mom_1m vs
rev_21d is −0.97 (they are the same number with opposite signs),
mom_12_1 vs mom_accel is −0.94, lowvol_60 vs downside_dev is +0.97.
No signal has a measurable edge
| signals at |t| ≥ 2 | 0 (1.6 expected by chance at 5%) |
| best |t| | 1.55 (downside_dev) |
| signs matching their a-priori prior | 13/31 — binomial p = 0.24, i.e. a coin flip |
| mean individual |IC| | 0.0140 |
| mean individual IC (signed) | −0.0058 |
t-stats are deflated by √(horizon/stride) for overlapping forward windows.
Without that correction they would be twice as large, which is the difference
between "nothing here" and two apparent discoveries.