With seven archived releases, calendar means raised US industry return forecast error by 5.8 percent in 2016–2026

We tested forecasts of relative monthly returns for 30 US industries from January 2016 to August 2026. Seven archived public data releases supplied all model inputs. Each fit stayed fixed until the next archived release. A calendar forecast used 20 past returns for the target calendar month. The reference used all 240 past monthly returns. The calendar forecast had 5.80 percent more squared error. The conditional 95 percent time-block interval was 3.91 to 8.17 percent. A fixed mix with 25 percent calendar forecast also had more error. A zero relative-return forecast had the least error. A calendar portfolio with long exposure one and short exposure one lost 7.36 percent a year before costs and 12.29 percent after assumed costs. These are annual arithmetic means. Revised inputs changed the selected portfolio in 67 of 128 months but gave a similar forecast error increase. Earlier retained outcome labels also gave a similar result. Training data were 3 to 37 months old. The sparse archive does not test continuous monthly updates. A related negative result was known before this study. The result is limited to this fixed forecast protocol and historical period. The package gives code, source hashes, all forecasts, model states, results, figures and repeat instructions. Independent scalar checks passed. A fresh run matched all 16 numerical output files.

return forecasts · industry portfolios · calendar seasonality · point-in-time data · data revisions · chronological evaluation · negative result · paper trading