LLM-Powered Virtual Population for Demand Simulation and Pricing

The article presents a novel LLM-powered virtual population model designed for demand simulation and pricing decisions, leveraging rich unstructured product information such as text and images. The model utilizes a finite mixture of customer personas to derive persona-level purchase probabilities, which are aggregated to form a predictive distribution of aggregate demand. Tested on an H&M fashion dataset, the simulator outperforms traditional models in predictive accuracy, allowing practitioners to make informed pricing decisions that account for demand uncertainty and various pricing objectives.