Series: EPU, UMCSENT, NFCI, STLFSI4 | z(36m/156w) level, lag-1 release, PIT | SPY long/flat, 10bps/switch | IS<2016<=OOS | Bonferroni bar p<0.00625 over 8.
Forward-return IC (block-permutation p, block 5)
| series | horizon | n | corr | perm p | last |
|---|---|---|---|---|---|
| EPU | 1d | 5720 | +0.003 | 0.792 | 2026-09-01 |
| EPU | 21d | 5720 | +0.045 | 0.096 | 2026-09-01 |
| UMCSENT | 1d | 5720 | -0.005 | 0.707 | 2026-08-01 |
| UMCSENT | 21d | 5720 | -0.020 | 0.488 | 2026-08-01 |
| NFCI | 1d | 5720 | -0.019 | 0.127 | 2026-09-25 |
| NFCI | 21d | 5720 | -0.015 | 0.595 | 2026-09-25 |
| STLFSI4 | 1d | 5720 | +0.000 | 0.980 | 2026-09-18 |
| STLFSI4 | 21d | 5720 | +0.019 | 0.497 | 2026-09-18 |
Directional overlays (long/flat, matched-exposure placebo)
Adopt bar: placebo p < 0.00625.
| series | dir | scope | Sharpe | buy-hold | exposure | placebo p | survives |
|---|---|---|---|---|---|---|---|
| EPU | long z>0 | IS | 0.14 | 0.46 | 38% | 0.0000 | no |
| EPU | long z>0 | OOS | 0.61 | 0.88 | 50% | 0.0000 | no |
| EPU | long z<0 | IS | 0.58 | 0.46 | 62% | 0.0000 | no |
| EPU | long z<0 | OOS | 0.60 | 0.88 | 50% | 0.0000 | no |
| UMCSENT | long z>0 | IS | 0.78 | 0.46 | 63% | 0.0000 | no |
| UMCSENT | long z>0 | OOS | 0.28 | 0.88 | 47% | 0.0015 | no |
| UMCSENT | long z<0 | IS | 0.03 | 0.46 | 37% | 0.0015 | no |
| UMCSENT | long z<0 | OOS | 0.92 | 0.88 | 53% | 0.0000 | no |
| NFCI | long z>0 | IS | 0.05 | 0.46 | 43% | 0.0000 | no |
| NFCI | long z>0 | OOS | 0.30 | 0.88 | 39% | 0.0005 | no |
| NFCI | long z<0 | IS | 0.70 | 0.46 | 57% | 0.0000 | yes |
| NFCI | long z<0 | OOS | 1.01 | 0.88 | 61% | 0.0000 | yes |
| STLFSI4 | long z>0 | IS | -0.02 | 0.46 | 29% | 0.0035 | no |
| STLFSI4 | long z>0 | OOS | 0.78 | 0.88 | 35% | 0.0000 | no |
| STLFSI4 | long z<0 | IS | 0.74 | 0.46 | 71% | 0.0000 | no |
| STLFSI4 | long z<0 | OOS | 0.32 | 0.88 | 65% | 0.0115 | no |
Mode-2 controls (is it just a calm-regime / momentum proxy?)
| series | corr(VIX z) | corr(mom21) | IC raw | IC partial |
|---|---|---|---|---|
| EPU | +0.25 | -0.01 | +0.040 | +0.042 |
| UMCSENT | -0.13 | -0.04 | -0.013 | -0.016 |
| NFCI | +0.38 | -0.23 | -0.022 | -0.041 |
| STLFSI4 | +0.45 | -0.29 | +0.011 | -0.009 |
Survivor robustness (NFCI z<0) vs a plain VIX overlay
| overlay | scope | Sharpe | buy-hold | exposure |
|---|---|---|---|---|
| VIX z<0 (comparator, lag-1) | IS | 0.21 | 0.46 | 59% |
| VIX z<0 (comparator, lag-1) | OOS | 0.86 | 0.88 | 63% |
| NFCI z<0 (lag-2) | IS | 0.71 | 0.46 | 57% |
| NFCI z<0 (lag-2) | OOS | 0.85 | 0.88 | 61% |
NFCI z<0 yearly net Sharpe: 2004:0.97; 2005:-0.69; 2006:0.00; 2007:-0.39; 2008:0.00; 2009:1.32; 2010:0.87; 2011:0.20; 2012:1.24; 2013:2.58; 2014:1.18; 2015:0.00; 2016:0.00; 2017:2.32; 2018:0.33; 2019:1.65; 2020:1.23; 2021:2.01; 2022:-0.85; 2023:0.00; 2024:1.64; 2025:0.94; 2026:1.37
Verdict
REJECT — ['NFCI:long z<0'] clears the pre-registered placebo bar in both IS and OOS, but the survivor (NFCI long-z<0) is a market-derived financial-conditions (calm-regime) composite: at the later lag-2 release it no longer beats buy-hold (see table), it correlates +0.38 with VIX, and its yearly Sharpe is 0.00 (flat) in 5 years and negative in 3. That is the already-rejected calm-regime family, not a new exogenous mechanism. EPU / UMCSENT / STLFSI carry no SPY timing.
Reading
- Level z, both directions, so an inverted mechanism (stress precedes bounces) is tested too. Sign alone is not a verdict.
- EPU is news-text-derived and UMCSENT a survey — genuinely exogenous to the price series, unlike VIX/credit spreads. Runtime 29s.