Granularity of weighted averages and use of rate statistics in AggPro

Abstract

AggPro predicts baseball statistics by utilizing a weighted average of predictions provided by several other statistics projection systems. The aggregate projection that is generated is more accurate than any of the constituent systems individually. We explored the granularity at which weights should be assigned by considering four possibilities: a single weight for each projection system, one weight per category per system, one weight per player per system, and one weight per player per category per system. We found that assigning one weight per category per system provides better results than the other options. Additionally, we projected raw statistics directly and compared the results to projecting rate statistics scaled by predicted player usage. We found that predicting rate statistics and scaling by predicted player usage produces better results. We also discuss implementation challenges that we faced in producing the AggPro projections.