Application of discrete choice models to choice-based revenue management problems: A cautionary note

Choice‐based revenue management algorithms directly integrate a discrete choice model of customer behavior into the optimization function. In the classic formulation by Talluri and van Ryzin (TvR), the choice set is defined to include a no‐purchase alternative, and the utility of this alternative is set to zero ‘without loss of generality’. We show that only a limited number of discrete choice models can be used with the TvR formulation, and that misuse of the TvR formulation can result in significant estimation and forecast errors. Based on the Matching Shares Property of multinomial logit (MNL) models, we show that estimated choice model parameters will match probabilities in the estimation data set for MNL models that include alternative‐specific constants (ASCs). MNL models that exclude ASCs may be able to match probabilities, but in this case the analyst cannot blindly set the no‐purchase utility to zero. This property holds regardless of which method is used to estimate choice model parameters.