DATA-DRIVEN TWO-STAGE INVENTORY PROBLEM

Date/Time
Date(s) - 15/09/2021
11:00 - 13:00

Categories No Categories


SEMINAR

Topic: AYÇA SİMAY ÇOLAK

Time: Sep 15, 2021 11:00 AM Istanbul

Join Zoom Meeting

https://zoom.us/j/6547746234?pwd=ZENZNWtCbUlQRjVMMVFneWtxZGlzZz09

In this thesis, we consider two-stage newsvendor problem where the decision maker selling a seasonal product only uses the historical demand information in her decisions. In our setting, there are two decisions to be made: the order quantity, and a marked-down price. We decide on how many products to order for the first stage, as well as how to set a marked-down price for remaining unsold inventory in the second stage. To solve the problem considered, data-driven models which do not require any distributional assumption are provided. Specifically, we propose six data-driven methods that solve the problem hierarchically in addition to another method which finds the order quantity and the marked-down price for the remaining inventory simultaneously by using a mixed integer linear program. We conduct a numerical study to evaluate the performance of proposed models. To do so, we generate the data from selected demand distributions and divide it into a training data and a testing data. The generated data is a function of the way that decision were made historically. We make a definition of the relevancy level based on what decisions the data depends on and address the effect of the data relevance numerically. Also, we determine the importance of size of the training data with the numerical study. For some given demand distributions, we find exact results to use in the performance comparison of proposed models. Then, we measure the percentage differences between the average profits obtained from models and the expected values obtained by assuming the known demand distributions. Numerical experiments on the generated data shows that our data-driven models perform well and their average profits are close to the exact results achieved under the assumption that the demand distributions are known, especially when larger training data size is used. Lastly, we measure how many times one model is the best among testing samples and compare models based on their performances.