Behavioral Neural Networks

More on Shaowei Ke

Abstract: We provide an axiomatic foundation for a class of neural network models applied to decision making under risk, called the neural-network expected utility (NEU) model. We weaken the independence axiom from expected utility theory in a new way consistent with experimental findings. We show how to use simple neurons, referred to as behavioral neurons, in the NEU model to capture behavioral biases such as reference dependence and the certainty effect. Empirically, we find that using standard estimation approaches, the NEU model has small training error but large testing error partly due to overfitting. However, by requiring the neural network to carry behavioral neurons, the errors can be significantly reduced.