Consumer Demand Modeling in a Grocery Delivery Service With Adaptive Pricing Mechanisms

This scientific research is dedicated to investigating dynamic pricing strategies in the context of same-day delivery of goods, taking into account user behavior. The goal was to compare various pricing approaches and find the most optimal strategy based on simulation analysis, customer behavior modeling, and reinforcement learning. To achieve this goal, the problem was formalized as a Markov decision process, and a simulator with stochastic customer request times and travel times between points was built. Within this simulator, a fast routing algorithm was implemented, considering pickup and delivery constraints, time windows, and vehicle capacity. To simulate user behavior and choice, a multinomial logit model, calibrated on real data from Sbermarket company, was constructed. Linear models and gradient boosting-based models were trained to approximate the value function of future states in the Markov process using reinforcement learning methods. Simulation results showed that advanced dynamic pricing strategies lead to increased revenue and served customers compared to optimal static pricing. The optimization strategy based on gradient boosting demonstrated the best results, resulting in a 4.3% revenue increase and a 12% increase in served customers. This study emphasizes the advantages of dynamic pricing and suggests further development directions, such as utilizing neural network models based on route plan embeddings for even more effective price policy management.