Misadventures in Monte Carlo

Abstract

Estimating probability is the very core of forecasting. Increasing computing power has enabled researchers to design highly intractable probability models, such that model results are identified through the Monte Carlo method of repeated stochastic simulation. However, confidence in the Monte Carlo identification of the model can be mistaken for accuracy in the underlying model itself. This paper describes simulations in a problem space of topical interest: basketball season forecasting. Monte Carlo simulations are widely used in sports forecasting, since the multitude of possibilities makes direct calculation of playoff probabilities infeasible. Error correlation across games requires due care, as demonstrated with a realistic multilevel basketball model, similar to some in use today. The model is built separately for each of 20 NBA seasons, modeling team strength as a composition of player strength and player allocation of minutes, while also incorporating team persistent effects. Each season is evaluated out-of-time, collectively demonstrating systematic and substantial overconfidence in playoff probabilities, which can be eliminated by incorporating error correlation. This paper focuses on clarifying the use of Monte Carlo simulations for probability calculations in sports.