Logit neural-network utility

Abstract

We introduce stochastic choice models that feature neural networks, one of which is called the logit neural-network utility (NU) model. We show how to use simple neurons, referred to as behavioral neurons, to capture behavioral effects, such as the certainty effect and reference dependence. We find that simple logit NU models with natural interpretation provide better out-of-sample predictions than expected utility theory and cumulative prospect theory, especially for choice problems that involve lotteries with both positive and negative prizes. We also find that the use of behavioral neurons mitigates overfitting and significantly improves our models’ performance, consistent with numerous successes in introducing useful inductive biases in the machine-learning literature.