Forecasting amidst Covid19 like events — How we can support global retailers like Kaufland or Lidl
by Arpit Jain (Arpit.Jain@sas.com) & Johannes Trummer (Johannes.Trummer@sas.com)
Since the declaration of Covid19 as a pandemic, a considerable impact is seen both in our personal as well as professional life. For example, social distancing, restricted livings and lock-downs have forced people to adapt to new ways of living. One of the major impact is seen in the shopping behavior of individuals — transition from mortar and brick stores to online channels, panic buying, pantry loading etc. This changing consumer demand pattern strongly influences the daily work of retail from purchasing, demand planning to restocking the shelves. Retailers or people, who are ensuring to meet the demand of groceries, are new found heroes for the general masses (https://www.linkedin.com/posts/kaufland-deutschland_schwarzgruppe-gemeinsamstark-schwarzdl-activity-6650361207675068416-vUXq).
In this post, we describe how we — Arpit Jain and Johannes Trummer (Customer Advisors at SAS) are analyzing these challenges, especially in context to demand forecasting, and how retailers like Kaufland, Lidl, Rossmann, Rewe etc. can benefit with our strategies. At the same time, we would also invite you all to an open discussion about further improvements and expansions to other suggested methodologies.
Crazy demand for daily products
An analysis of the Federal Statistical Office in Germany shows the demand for daily products — beer (Bier), disinfectants (Desinfektionsmittel), flour (Mehl), soap (Seife) and toilet paper (Toilettenpapier). The actual demand for the products on August 5, 2019 is set as the reference point of 100% and accordingly, shows how the demand has evolved ever since. We observe an absurd eight-fold demand growth for disinfectants and likewise, a steady exponential increase in the demand for soap and toilet paper in March 2020. This may be the result of the panic buying or pantry loading behavior by individuals in lieu of the restricted living announcement made by German government. Given such extraordinary demand patterns, the question that most retailers are now facing is “How can meaningful and accurate short-term and long-term demand forecasts be generated?”
To solve this challenge, we created demand forecasts based on the given data from www.destatis.de using SAS Visual Forecasting solution. We forecast using both conventional time series ARIMA modeling and modern machine learning modeling — Panel Series Neural Network and a Stacked Model, which combines both time series and Neural Network models. As additional events, we included the start date and duration of restricted living/curfews into our models.
Applying general time series and machine learning models to all products
In order to identify the best possible approach to effectively handle short-term demand patterns, we experimented different modeling strategies. First, we applied the same forecast models with tuned parameters to all five products.
The results show that the recent demand patterns have not been pragmatically handled by the forecasting models. Although not the best, but for most cases the machine learning models outperform conventional time series models like ARIMA. We see this more clearly by comparing the MAPE (Mean Absolute Percent Error) scores between models, the standard most-widely used KPI measure for forecasting. We observe an exorbitant increase in the general trend of the future demand for all the products. For example, the demand for soap is forecast to rise above 3,000% by the end of April 2020, which seems unrealistic compared to the 100% starting point in August 2019. Similar patterns are also observed for other products — disinfectants with a maximum of over 3,500%, flour with a maximum of approx. 1,300% and toilet paper with a maximum of approx. 1,700%. Noticeably, for disinfectants and flour, a saw-tooth pattern is seen, which can be explained by two different causes. On the one hand, the unexpected and extremely high demand for disinfectants led to under supply or out of stock conditions. All stocks were used up and no disinfectant could be sold, which is also visible in the demand pattern. On the other hand, the formation of the saw-tooth pattern is intensified by the Covid19 event. We explain the general upward trend in the forecast for the individual products by the above-average increase in demand for various products in general, or rather, as the basis of the forecast is a comprehensive view of general consumption, triggered by the lock-down situations.
This leads us to look at the respective products in detail and thus to forecast based on product segmentation. For this purpose, the information on past demand is already divided and the forecasts for the respective products are created individually. In our follow-up post “Forecasting amidst Covid19 — Improving Forecasts with Segmentation and Pandemic Model” we will describe the enhancements of applying segmentation and including a pandemic model as an indicator to add value to the forecast. Furthermore, we will discuss our results and limitations we face while setting up the different models.
