Session hosted by: Dr Cátia Silva, Senior Data Scientist, bol.com
What was the session about?
In this talk, Dr Silva highlighted the challenges experienced by a forecasting data science powered team due to the pandemic and how bol.com implemented solutions to predict the unpredictable. Predicting the unpredictable is not easy, especially during the pandemic where the retail industry saw a lot of changes and needed to ensure they are forecasting accurately to save money. The pandemic led to consumers needing more of certain products than others. A well-known example is a sanitiser, consumers never really needed sanitiser in such quantities, but it became a high demand product. Additionally, the pandemic led to an increasing online demand due to physical stores being closed.
The effects of COVID-19 created a change in the general patterns of the retail industry and data forecasting which became unreliable. Usually, sales forecasting is based on informative factors such as past sales, seasonality, national holidays, promotions, and campaigns. But this became very unreliable due to the pandemic. It became necessary to identify corona influenced sales and create time features to represent important moments, however, this would not be a simple task with hundreds and thousands of products.
At bol.com the Corona severity feature was implemented to see how severely corona was affecting sales to tackle the problem. The severity feature was based on the news and press conferences that determine higher or lower demand based on the increase or decrease in the restrictions placed by the government.
Opinion & Key Takeaways:
This session is truly relevant to the current global environment and probably will still be relevant until things stabilize for the long run. It is something you would not really think of as a consumer but looking at the business perspective you can really see how unpredictable and challenging forecasting becomes with such disruptions.
Bol.com’s use of modelling with the corona severity feature led to better accuracy in the forecasting. As you can see on the diagram where there were large gaps in the forecast vs. actuals in the beginning, to then visibly closer lines depicting the 30% relative improvement on the validation set. Retail companies should investigate solutions like the Corona Severity Feature to assist them in staying up to date and as accurate as possible during unpredictable times like now.