CORAL Seminar: Johan Clausen, Aarhus University

Title: Emperical Risk Minimization and Conditional Inventory Optimization

2020.05.25 | Solveig Nygaard Sørensen

Date Thu 04 Jun
Time 14:00 15:00
Location Online via Zoom

PhD Seminar

Presenter: Johan Clausen


Imagine you are the operational VP of a national bakery chain and you wish to expand to new locations in your country. Your data analysts tell you that the most important features that determines how much of your products you sell is a set of socio-demographic variables collected on the residents living close to the bakery location. Moreover, based on your existing bakeries you can estimate the conditional relationship between socio-demographic variables and bakery sales with a satisfactory degree of accuracy and you can expect that the estimated relationships are valid across the country. However, looking at the new possible locations you observe, that the socio-demographic profile of the new locations are very different from your existing stores. Therefore, you know that projecting daily inventory levels from your existing stores when determining the daily inventory of your new potential stores would likely lead to either a large amount of lost sales or a large amount products needing to be disposed of. As a person educated in economics you know that you can do conditional forecasting, but using basic methods only gives you a conditional mean estimate of sales. The question you are left with is then, how do I take lost sales costs and disposable costs into account when determining the inventory level at the new store locations?

Exactly this type of newsvendor problem is solved in the pioneering paper “The Big Data Newsvendor: Practical Insights from Machine Learning” by Ban and Rudin (2019) that uses the principle of empirical risk minimization to formulate and solve a data driven conditional newsvendor problem. Subsequent papers have expanded upon and improved on the work of Ban and Rudin (2019), but all subsequent papers have only examined the newsvendor problem. This leaves a large research question open; can the principle of empirical risk minimization be used to solve other conditional inventory problems? 

I am at the moment in the process of writing a paper that seeks to both clarify the findings of the literature on the data driven conditional newsvendor problem, and show how one can solve other data driven conditional inventory problems using the ideas introduced in this new field of literature. 

I will in this seminar therefore introduce how the principal of empirical risk minimization is used in the literature to solve the conditional newsvendor problem, how I use the ideas introduced to solve an order-up to inventory problem, and present some considerations on limitations and future applications of the this new field of research.

Organised by Professor Sanne Wøhlk

CORAL seminars