How predictive performance varies with data splitting.

<p>RMSE (<b>A</b>.) and correlation (<b>B</b>.) for time series forecasts of play in 30 game structures, varying folds in cross-validation from 30 to 2. The full model consistently has lower prediction error and higher correlation than the baseline model and the fEWA model until there are only two folds. It is, in general, difficult to make accurate predictions when the ratio of observational units to folds is small. In the case of predicting aggregate and dynamic play, the game structure itself is the observational unit, and we only have thirty, so it’s not surprising that performance can degrade at two folds depending on the particular random realization of fold assignments.