Five-year return means gave no clear forecast gain with annual US market archives in 2005–2025

We tested whether the mean US market excess return from the last five years improved monthly forecasts over the mean from all earlier years. Official annual archives set all historical inputs. Each file was assumed available by the end of September and supplied forecasts for the next October through September. The test had 240 months from October 2005 to September 2025. The five-year mean had 1.67 percent more squared error. Its out-of-sample R-squared was -1.67 percent, with a descriptive 95 percent time-block interval of -4.03 to +0.40 percent. A ten-year mean, a zero floor, a fixed mix and zero gave no clear gain. Revised values at the same input dates gave a similar result. With monthly updates to more recent revised inputs, the five-year score was also below the reference. In a market and cash simulation at an assumed five-basis-point cost, the five-year rule earned 5.42 percent mean excess return a year, against 5.92 percent for the expanding rule. The paired gain interval was -2.07 to +1.10 percentage points a year. The source labels contain later revisions. The annual input endpoints are 3 to 14 months old, and original release dates are assumed from provider records. The evidence supports a narrow negative benchmark, not reliable stock-return prediction. Code, source hashes, all forecasts, model states, cost paths, uncertainty draws, figures and independent checks are preserved. A November 2026 paper forecast remains unscored.

equity premium · return forecasts · historical mean · data archives · data revisions · chronological test · trading costs · negative result