Physicists Disrupting Retail - Part 2

In my last blog I wrote about how physicists are ideally suited to become the cornerstone of digital businesses and how the emerging field of Data Science gives us a new home in the private sector.
In this article I would like illustrate this with a real-world example: Replenishment of perishable and non-perishable goods, all products at a given supermarket store.

CERN, Physik, Machine Learning, Handel, Disposition

Replenishment in Retail

Academically, this spans a wide range of topics in Operations Research of which, optimal solutions have been found in the past 50 years for specific aspects in certain, sometimes idealized conditions or with dedicated requirements.

Reality is often much different: First of all, the line between perishable and non-perishable goods is rather blurry. Fresh bread is certainly perishable, a baguette will be stale in the evening, if not before. Similarly, convenience food, fresh meat, fruit and vegetables spoil quickly. But even something as basic as milk is quite diverse: there is fresh milk which keeps just a few days, micro-filtrated milk and then pasteurized long-life milk, not to mention the various yoghurts or young and mature cheeses you can find in a typical dairy section. On the other hand, frozen or canned food, drinks, tea and coffee have a long shelf life. A typical supermarket chain has around 3,500 perishable products alone and often more than 20,000 different products in their whole assortment range. In this environment, applying generalized models built for idealized conditions with dedicated requirements severely limits optimal decision making.

The key challenge to be addressed is how to keep sufficient stock levels of all the varying products to meet customer demand across several diverse locations. Customer demand is influenced by a myriad of factors, e.g. season, day of the week, weather, holidays, promotions, assortment range and many others. In addition to the already complex environment, new products can be introduced, others phased out, some limited to specific order cycles and lead times or in multiples of specific lot sizes. To complicate matters further, data are recorded by complex IT systems which leave ample opportunities for errors in the data such as incorrect stock levels.

In line with academic research, the most benefit can be obtained if this replenishment challenge is solved based on predicting the full probability density distribution of the future customer demand. In academia one can just say “assuming the distribution is known”, however, in a practical situation one has to develop a method to actually calculate the distribution. The calculation of this distribution is further complicated by the fact that the true customer demand is not observable in most realistic scenarios, only recorded sales data are available. In general, customer demand and observed sales are related but not identical, as products are swapped for alternatives (substitution) or sales are lost if the desired item is not available and no suitable alternative is found. Sometimes “lost sale” of a specific item leads to the customer abandoning the shopping altogether and “ill will” costs have to be considered as well.

Physicists become Data Scientists

Replenishment in retail is a challenging problem – but one for which physicists are ideally suited to solve: Due to our training and long experience, we’re used to thinking in terms of probabilities and distributions, rather than in deterministic sequences of events. This allows us to understand business needs on a wide scale. Our experience in developing sophisticated machine learning algorithms allows us to predict the individual probability density distribution of future demand from (censored) sales data for specific products at specific locations depending on all influencing factors.

A recent article jointly authored by McKinsey and Blue Yonder highlights the business advantage customers can obtain from the application of machine learning based generation of probability densities in a real-life setting.

Deep understanding of the field, solving complex statistical challenges and building sophisticated machine learning algorithms make up the natural habitat for physicists turned Data Scientists. This, combined with business experts and software engineering, operations, sales, marketing this is the recipe for the next generation of algorithmic enterprises.

Dr. Ulrich Kerzel Dr. Ulrich Kerzel

earned his PhD under Professor Dr Feindt at the US Fermi National Laboratory and at that time made a considerable contribution to core technology of NeuroBayes. He continued this work as a Research Fellow at CERN before he came to Blue Yonder as a Principal Data Scientist.