Combine and conquer: model averaging for out-of-distribution forecasting

Abstract

Travel behaviour modellers are increasingly interested in using models from outside the traditional choice modelling area, first incorporating ideas from behavioural economics, such as in regret modelling, before looking at mathematical psychology and machine learning. A key question arises as to how well these different models perform in prediction, especially when predicting trips of different characteristics from those used in estimation. This paper first compares the elasticities and model fit of different models, bringing together models as diverse as logit, random regret, decision field theory and neural networks. We highlight differences in elasticities and also note that the prediction performance deteriorates at different rates for different models when moving further away from the estimation data. We then develop a model averaging approach that allows us to make the most of the entire collection of models and estimate weights for different models as a function of distance away from the estimation sample. Keywords: choice modelling; forecasting; machine learning; mathematical psychology; mode choice; model averaging