Demand Estimation and Price Optimization for Substitutable Products

The advances in machine learning and data analytic tools have significantly facilitated the demand learning process, where companies estimate how customers will respond to various levels of prices. The demand forecasts then feed as inputs into the price optimization models to determine the best prices to meet their objectives. This process imposes several challenges in research and business practices. For the demand estimation stage, Machine learning models are widely used in the industry with outstanding performance in prediction accuracy, but they suffer from a lack of interpretability. Moreover, there is an increasing need to model substitution effects across products due to intensified levels of competition. This leads to high dimensional demand models that are likely to be intractable in optimization. The complexity in the demand model will translate to challenges in the optimization stage, where the model needs to address the high dimensional nonlinear forecast inputs and business constraints. In this project, we propose to develop and empirically verify a price optimization framework that captures product substitution patterns through flexible demand structures. Our preliminary studies develop a pairwise comparison estimation model to capture substitutions and prove its robustness under error term misspecifications. We plan to further investigate the estimator's theoretical properties and numerical performance through extensive numerical analysis and studies on real data. On the optimization side, we propose an approximation scheme by utilizing the intermediate variable to reduce the dimension of inputs. By discretizing the intermediate variable, we can also deal with the nonlinear inputs and business constraints. We plan to conduct numerical analysis using real data to identify the best type of intermediate variables. We also initiate an industry collaboration with Meituan bike, one of the largest bike-sharing companies in China. We plan to verify the proposed frameworks by utilizing real bike-sharing historical data, as well as designing and implementing field experiments. The research output of this project will contribute to the literature in the area of choice-based demand estimation and optimization. Moreover, we expect the project to have real business impacts on companies in their demand learning and pricing decision-making process.