Pricing Items Based on Their Own History: Transformer-Generated Deep Isotonic Pricing Networks

This paper introduces a dynamic pricing framework for e-commerce that leverages data science to optimize revenue. The framework employs a dual-module system combining demand forecasting and optimal pricing. Key innovations include 1)using a Transformer encoder to generate model weights based on historical data and 2)introducing a Double Weighted Mean Absolute Percentage Error to improve accuracy. Field experiments on 1688.com demonstrated increased platform profitability.