Past returns reduced errors in daily return size but not direction for five US indexes in CNNpred data, 2015–2017

We tested forecasts for daily direction and absolute price return in five US indexes. The source was the open CNNpred dataset. We selected model settings in 2013–2014 and tested 725 later sessions from January 2015 to November 2017. Each valid forecast used only earlier index returns. Models refitted each month. Own-history ridge reduced squared error for absolute returns by 21.9% against the expanding mean, with a 95% date-block interval of 15.2% to 28.9%. It also reduced error by 11.0% against the recent 20-session mean, with interval 3.5% to 17.2%. The Holm-adjusted primary p value was 0.0004. The global direction model increased log loss by 1.07%, with interval 0.37% to 1.80%. Its accuracy was 50.9%, against 53.0% for the historical-frequency reference. An extra session of input delay preserved the size gain. An invalid control with target-session inputs reached 91.8% accuracy. This shows how future information can inflate scores. It does not test the original paper's code. The package contains source data, provenance, code, forecasts, fit dates, checks, resampling draws, and figures. The source is a retrospective snapshot. Original historical releases and exact foreign closing timestamps remain unverified. The findings apply to these models and this period. We did not test trading profit.

daily index forecasts · absolute returns · direction prediction · chronological tests · information time · forecast baselines · negative results