Personalized Dynamic Pricing Strategies for Perishable Products: A Contextual Bandit Algorithm Approach

ABSTRACT: In fast-moving retail environments, pricing perishable items effectively is a complex task due to short shelf lives, variable demand, and inventory fluctuations. Dynamic pricing optimization is the part of supply chain management, allowing businesses to improve sales, reduce waste, and optimize inventory levels. This study applies contextual bandit algorithms: Thompson Sampling and LinUCB to develop adaptive pricing strategies that respond to customer behaviour and product perishability in real time. Using a cleaned grocery retail dataset, a simulation environment was created with engineered features including customer segments, days to expiry, and inventory levels. Five pricing strategies were evaluated: Thompson Sampling, LinUCB, Fixed Pricing, Random Pricing, and Greedy Pricing. Performance was assessed using cumulative reward, regret, sell-through rate, and conversion rates across segments. Thompson Sampling achieved the best results, with a cumulative reward of 77.4% and the lowest cumulative regret, demonstrating best adaptability and effective model. The results indicate that learning-based models significantly outperform static approaches and are well-suited for dynamic pricing of perishable goods in modern retail.