Shrinking ten-year industry means gave no clear relative-return forecast gain in 71 monthly archive tests, 2016–2026

We tested whether ten-year industry return means predict monthly returns relative to an equal-weight industry basket. Ten dated archives supplied fixed forecasts for thirty US industries. A full calendar-month buffer followed each capture month. Each forecast stayed fixed for up to twelve months. The test has 71 distinct target months in 2016–2026, with gaps. A simple moment rule reduced the past means toward zero. It had 0.698% more squared error than a zero relative-return forecast. Its error-reduction score was -0.698%, with a descriptive 95% interval from -1.868% to +0.165% across seven capture-year groups. The full and half past means also gave negative point scores. The moment rule had 1.017% less error than the full mean, but that interval included zero. Revised-data checks gave similar results. Most of the primary error increase came from the 2020 capture-year group. Sparse old inputs, seven groups, late labels, source changes and the simple noise estimate limit inference. No trading strategy was tested. Code, source hashes, all forecasts, states, uncertainty draws, negative findings, independent checks and an exact thirteen-file replay are preserved. Fixed forecasts for December 2026 to November 2027 remain unscored.

industry returns · return forecasts · mean estimation · shrinkage · dated archives · chronological evaluation · negative result