Leveraging Assortment Similarities for Data-driven Choice Predictions

Choice models are fundamental tools in revenue management, used for predicting customer demand and optimizing operational decisions such as pricing and assortment. Parametric models, like the MNL and nested logit, facilitate efficient estimation and decision-making but are prone to misspecification, which can adversely impact prediction accuracy. Nonparametric models like the rank-based model, while capturing complex substitution patterns and offering more accurate demand predictions, are computationally intensive, difficult to optimize, and require larger datasets to avoid overfitting. This trade-off complicates the selection of a suitable choice model in practice. In this work, we propose a fully data-driven framework that combines the strengths of both parametric and nonparametric approaches. Given access to historical transaction data, our method predicts choice probabilities for a given test assortment by leveraging data from historical assortments that are most ``similar'' to the test assortment, taking inspiration from nearest neighbor methods in machine learning. Our approach is model-free, requiring no estimation of any model parameters. Moreover, we show that our method provides universal consistency of both the demand predictions and ensuing assortment decisions, that is, they converge to the true values as the size of the historical data increases, regardless of the underlying choice model generating the data. Extensive numerical experiments on synthetic data demonstrate the superior prediction and decision performance of our proposed framework compared to parametric choice models like MNL and LC-MNL. Additionally, we introduce a hybrid approach that combines predictions from our similarity-based method with those obtained from parametric models, which enhances prediction accuracy even in scenarios with limited transaction data. This integration bridges the strengths of both approaches, making our framework practical and robust in real-world settings.