Probabilistic forecasting

why model imperfection is a poison pill

Roman Frigg

pp. 479-491

in: Hanne Andersen, Dennis Dieks, Thomas Uebel, Wenceslao J. Gonzalez, Gregory Wheeler (eds), New challenges to philosophy of science, Berlin, Springer, 2013

Abstract

Foretelling the future is an age-old human desire. Among the methods to pursue this goal mathematical modelling has gained prominence. Many mathematical models promise to make probabilistic forecasts. This raises the question of exactly what these models deliver: can they provide the results as advertised? The aim of this paper is to urge some caution. Using the example of the logistic map, we argue that if a model is non-linear and if there is only the slightest model imperfection, then treating model outputs as decision relevant probabilistic forecasts can be seriously misleading. This casts doubt on the trustworthiness of model results. This is nothing short of a methodological disaster: probabilistic forecasts are used in many places all the time and the realisation that probabilistic forecasts cannot be trusted pulls the rug from underneath many modelling endeavours.