Modelling choices: The conflict between realism and prediction performance

Research in fields as diverse as mathematical psychology, choice modelling and machine learning has resulted in a range of methodologies for analysing individual decision-making with many similarities, but often with little to no spillovers or dissemination among them. These methods are of great relevance in travel behaviour research, but analysts often lack clear guidelines on which approach to use based on the context. The current study compares a range of models from the aforementioned fields focussing on out-of-sample prediction performance and applicability to policy making. Results indicate that machine learning algorithms might outperform other methods in terms of prediction accuracy on the estimation/training sets, but those benefits are diminished when applied to the hold-out/test sets. Contrary to that, choice models and psychological modelling specifications show higher stability between train and test sets, with the context impacting which model performs better. Mathematical Psychology models incur substantial computational costs while RUM retains clearer economic interpretation. Finally, demand elasticities/marginal effects tend to be more behaviourally realistic in choice models and mathematical psychology structures compared to machine learning, making it more difficult to apply the latter in real-world policy making.