Research On Dynamic Pricing Algorithm Of Perishable Goods Based On Demand Learning

Because perishable products are different from most daily necessities,for the sale of perishable products,first of all,their sales cycle is limited,while daily necessities have almost no sales period restrictions;secondly,due to the randomness of demand for perishable products,it is difficult to carry out Data collection to continuously update the demand distribution to achieve as soon as possible benefits.To sum up,for sellers to sell perishable products,it is very necessary to learn how to obtain the demand function that can achieve the maximum profit as soon as possible.Consumer demand will be affected by factors such as price adjustment,so sellers can use dynamic pricing technology to conduct demand learning on products with limited inventory within a limited time,which can lay a good foundation for better returns in the future.This paper studies a single product with limited inventory level and sales period.Because the demand for perishable products is random,it is impossible to know the functional relationship between product demand and price determination.Therefore,in order to simulate the customer’s purchase process for the product,it is assumed that the customer’s purchase volume for the product obeys a random number of Poisson distribution with the mean value of a given potential demand rate.This paper designs a dynamic pricing algorithm for perishable products based on demand learning,updates the test price range in each learning stage through the estimated optimal price,and estimates the parameters of the demand rate function,so that the final maximum yield is close to the true value of "God’s perspective" under the theoretical maximum yield.Then,based on the principle of the original algorithm,the shortcomings of the original algorithm were analyzed,and the algorithm was improved from four aspects: increasing the number of test prices in some learning stages,changing the price interval,repeating the test for the same price,and optimizing the termination conditions of the algorithm.In the empirical research,the original algorithm is studied in a simulated scenario: when learning the needs of customers with different price sensitivities,the algorithm gives the corresponding initial test price interval and price interval,and obtains the maximum return rate learned in the termination stage.The difference between the value and the theoretical value under "God’s perspective" is small.When different methods are used to estimate the parameters of a demand rate function in an exponential form,the algorithm results obtained by taking the logarithm of the demand rate function first and then using the least squares estimation method are similar to those obtained when using the nonlinear least squares estimation directly.A simulation scenario study of the improved algorithm:by learning the same demand rate function and given the same initialization conditions,the probability that the algorithm contains the optimal theoretical value of price at each stage is obtained after the algorithm is improved in any aspect.In order to increase,and it is guaranteed that the estimated value of the maximum rate of return will continue to increase in the subsequent learning,and further compared with the return value obtained under the original algorithm,the value of the improved algorithm has dropped significantly.Finally,by considering the changes of the initial inventory level of the product used for learning and the potential demand rate function with the magnitude,it is obtained that with the continuous increase of the magnitude,the final learned regression value of the algorithm continues to decrease,and gradually converges to 0,which proves the convergence of the original algorithm and the improved algorithm to a certain extent.Further through the comparison with other related algorithms,the effectiveness of the improved algorithm in this paper is confirmed,and it is proved that for perishable products,it is necessary to conduct demand learning in a limited sales cycle.Finally,the robustness of the proposed algorithm is shown by sensitivity analysis.