Archived seasonal industry forecasts lost 10.9% a year after assumed costs across five 12-month tests

We tested a seasonal return forecast for 30 US industry portfolios. Five archived files set all forecasts before each later 12-month test. The tests contain 60 months from 2016 to 2025. Forecasts stayed fixed during each test. The seasonal rule uses 20 earlier returns for the target calendar month. Its baseline uses each industry's last 240 monthly returns. Archived seasonal forecasts had 1.55% more squared error than this baseline. A paper portfolio bought the top six industries and sold short the bottom six. Its gross arithmetic mean was -6.41% a year. Its net mean was -10.87%. Assumed costs were 10 bp per dollar traded and 50 bp a year for short borrow. The five-block bootstrap interval was [-18.31, -3.44]% a year. A small-sample t interval was [-22.52, 0.77]%. Thus, inference depends on the method. Later data changed 31 of 60 monthly portfolios and 74 of 1,800 position assignments. Both forecasts used the same historical input dates. The later data gave a net mean of -9.97% a year. The paired change was +0.91 percentage points a year, with a block interval of [-2.79, 3.96]. Both versions kept a mean loss. The data contrast includes historical revisions and the CRSP return-method change. The sample is small, archive dates are selected, and costs are assumed. This is a historical reconstruction, not a prospective experiment. The package preserves code, input hashes, archive dates, forecasts, weights, scores, costs, resampling draws, and checks. Raw provider files remain in separate retained storage under their original terms.

industry returns · return seasonality · point-in-time data · historical data revisions · chronological tests · paper trading · negative result