A 50% mix with common correlations cut joint loss but gave no clear paper risk gain for 30 US industries in 66 months, 2015–2025

We tested fixed monthly forecasts for 30 US industry portfolios with seven dated input archives. The test has 66 distinct months from December 2015 to August 2025, with gaps. Each fit uses 120 source months. A 50% mix of sample covariance and a common-correlation matrix cut joint Gaussian loss by 5.1265 nats per return vector against sample covariance. The descriptive 95% interval from six capture-year groups was 1.7918 to 8.4208 nats. A separate sign-flip test gave p = 0.0625. Almost all the gain came from conditional industry contrasts in an exploratory score split. The Gaussian 90% region covered only 44 of 66 outcomes. At an assumed fee of 10 bp per dollar traded, the mix gave no clear risk gain in a long-only paper allocation with a 20% industry cap. Monthly standard deviation was 3.1502%, against 3.1410% for sample covariance. The reduction interval was -0.0684 to +0.0524 percentage points. Sparse old inputs, six groups and labels that are not proven first releases limit inference. We preserve code, source hashes, full forecast matrices, weights, scores, uncertainty draws, negative findings and 60 unscored future forecasts. We do not claim a new shrinkage method or reliable trading profit.

covariance shrinkage · industry portfolios · dated archives · risk forecasts · chronological evaluation · paper allocation · joint log score · reproducible research