Weather models improved catch allocation scores in 17 of 18 UK moth site-period tests in 1993-2015

We tested weather prediction across sites and periods with 29,041 moth trap nights at seven English and two Scottish ECN sites in 1993-2003 and 2004-2015. Each model trained only at other sites in the opposite period. A full weather model improved catch-weighted log scores over uniform weekly allocation in 17 of 18 site-period tests. Equal-year gains were 0.0313 to 0.0745 nats per caught moth in four region-period samples. Conditional four-test intervals were above zero in three samples and crossed zero in later Scotland. Later Sourhope had a negative score with 203 nights. That failure persisted with positive catches and three-day groups. These predictions condition on observed weekly totals and recorded nights. They do not estimate detection alone, population trends, pollution effects or causal climate effects. Few sites, missing zero reports and training uncertainty limit transfer claims. Code, aggregate scores, source hashes, retrieval routes and numerical checks are included.