# Dynamic Pricing on E-commerce Platform with Deep Reinforcement Learning

**Authors:** Jiaxi Liu, Yidong Zhang, Xiaoqing Wang, Yuming Deng, Xingyu Wu  
**arXiv:** [1912.02572](https://arxiv.org/abs/1912.02572)  
**Date:** 2019-12-05  
**Summary page:** https://www.summarizepaper.com/en/paper/1912.02572v1/

## Key points

- Authors introduce a framework for dynamic pricing in E-commerce using deep reinforcement learning (DRL)
- Model dynamic pricing as a Markov Decision Process (MDP) with four groups of business data representing different states
- Three key contributions compared to existing DRL-based dynamic pricing algorithms:
- Extend discrete set problem to continuous price set for more flexibility and accuracy
- Introduce novel metric called Difference of Revenue Conversion Rates (DRCR) as reward function
- Address cold-start issue of MDP through pre-training with historical sales data
- Offline assessments on real datasets from Alibaba Inc. and online field experiments on Tmall.com show superiority of DRCR over traditional revenue-based metrics
- Continuous price sets outperform discrete sets in extensive field experiments on 1000 stock keeping units (SKUs)
- Framework surpasses manual pricing strategies implemented by operational experts, showcasing superior performance in online retail environments

## Summary

In their paper titled "Dynamic Pricing on E-commerce Platform with Deep Reinforcement Learning," authors Jiaxi Liu, Yidong Zhang, Xiaoqing Wang, Yuming Deng, and Xingyu Wu introduce an innovative framework for addressing dynamic pricing challenges in E-commerce using deep reinforcement learning (DRL) techniques. The study focuses on modeling the dynamic pricing problem as a Markov Decision Process (MDP) by utilizing four distinct groups of business data to represent different states during each time period. The researchers propose three key contributions compared to existing DRL-based dynamic pricing algorithms. Firstly, they extend the traditional discrete set problem to a continuous price set, enhancing the flexibility and accuracy of pricing decisions. Secondly, instead of directly using revenue as the reward function, they introduce a novel metric called the difference of revenue conversion rates (DRCR), which proves to be more effective in optimizing pricing strategies. Thirdly, they address the cold-start issue of MDP by pre-training and evaluating models with carefully selected historical sales data. To evaluate their approach, the team conducts offline assessments using real datasets from Alibaba Inc. and online field experiments on Tmall.com, a prominent online shopping platform owned by Alibaba Inc. The results indicate that DRCR outperforms traditional revenue-based metrics commonly used in literature. Furthermore, extensive field experiments conducted over several months on 1000 stock keeping units (SKUs) demonstrate that continuous price sets yield superior performance compared to discrete sets. Ultimately, the researchers show that their framework significantly surpasses manual pricing strategies implemented by operational experts. Overall, this study provides valuable insights into leveraging deep reinforcement learning for dynamic pricing optimization in E-commerce settings and highlights the importance of innovative reward functions and continuous price sets for achieving superior performance in online retail environments.