Improving Australian Football League player performance forecasts using optimized nonlinear smoothing

This research demonstrates how exponentially-smoothed, one-step forecasts of Australian Football League (AFL) player performance data are improved by first applying a nonlinear (Tukey) smoother to the raw data. The player performance data are derived from a simplistic linear model, such as the one seen in an AFL ''fantasy'' football league. A smoothing macro allows experimentation with various running median combinations which are designed to eliminate the noise from each player's season data. Performing optimizations on each player's running median sequence in conjunction with the exponential smoothing parameter results in a noticeably lower mean squared error per player than either mean projection or simple exponential smoothing. A Monte Carlo simulation of the median sequence and smoothing parameter combinations creates confidence intervals for assessing the forecasts. The results are demonstrated on both a season and a match-by-match basis.