Bol.com is a retail platform with over 12 million prospects within the Netherlands and Belgium. To serve the wants of those prospects, it’s important that our staff gives correct forecasts to empower enterprise selections on a day after day foundation. With the outbreak of the COVID-19 pandemic, on-line purchasing behaviour went by way of a significant shift and demand for important and non-essential merchandise swiftly elevated (see cbs).

After months of battling producing any affordable forecasts, we managed to design a function which is ready to describe the dynamic modifications related to the COVID-19 pandemic. The addition of this function permits us to offer dependable forecasts within the brief time period and roll-out state of affairs forecasting for the long run, supporting totally different domains throughout the enterprise. Our strategy is definitely interpretable and explainable to stakeholders, main to higher information pushed selections.

In our forecasting panorama, we offer several types of gross sales forecasting, each on whole stage and likewise on product stage. These forecasts are then used on their very own for various functions throughout the organisation but additionally utilized by our staff as the primary drivers for operational planning forecasting. The totally different forecasts depend on totally different time-series modelling strategies, from linear fashions to gradient boosting algorithms. As such, it was vital for us to have an one-size-fits-all resolution which might scale throughout the totally different modelling approaches.

We began with an in depth information evaluation and studying on the data relating to COVID-19 restrictions to comprise the unfold of the virus. With this info we developed a severity index that translated the impression of the pandemic on our gross sales patterns. Our severity index ranges between 0 and 12, the place a 0 represents no COVID-19 associated restrictions and a 12 represents the tightest restrictions we encountered through the pandemic. Determine 1. reveals a illustration of the COVID-19 severity index.

This COVID-19 severity index is just like the publicly obtainable COVID-19 stringency index (obtainable right here). The primary distinction is that our severity index is tailor-made to the actual dynamics of the affect of the pandemic on our gross sales information, which isn’t simply depending on the kind of restrictions in place. For instance, the implementation of obligatory mouth masks in public transport, or the momentary closing of bodily outlets has the next impact on our gross sales sample than what the already obtainable stringency index would counsel. Creating our severity index additionally permits us to extra simply translate the brand new info on restrictions from press-conferences right into a future trying index, which is key for forecasting. As well as, we’re in a position to create higher and decrease bounds for the severity index based mostly on the earlier months information and with this allow state of affairs forecasting.

For the reason that COVID-19 severity index is only a time-series function, it’s simple to implement it throughout the totally different modelling approaches in our panorama. The outcomes of including the severity index to our fashions had been spectacular, the place we noticed as much as 40% relative enchancment on the imply absolute error of the validation interval (from March 2020 to Jan 2021). On high of that, the ahead trying forecasts additionally started to be far more according to expectations, which gave our stakeholders a significant forecast accounting for the most recent pandemic dynamics and in consequence additionally elevated their belief in our predictions.

For long run planning, we additionally aided our stakeholders by offering totally different state of affairs forecasts, each based mostly on a distinct doable final result of the pandemic. On this course of, stakeholders might request a state of affairs based mostly on a reference interval for which we’d then lengthen the severity index utilizing the identical values of that interval and forecast the gross sales accordingly, see Determine 2.

Throughout this era, we labored in even nearer collaboration with our stakeholders than earlier than, ensuring to take their skilled data under consideration to determine on the long run values for the severity index. The truth that the severity index is definitely comprehensible makes this strategy very clear which resulted in a swift adoption of the improved forecasts.

From provide chain to logistics operations and customer support planning, our resolution to forecasting demand throughout a pandemic has allowed us to offer dependable forecasts to empower enterprise selections within the brief time period, and roll-out state of affairs forecasting for the long run. We could not know but what the way forward for the pandemic seems like, however we are actually significantly better ready.

*This work was developed as a joint effort from Workforce Forecasting at bol.com by Asparuh Hristov, Bjarnthor Egilsson, Cátia Silva, Erik Mulder , Eryk Lewinson, Roberto Carcangiu, Susanne Tak, Tavis Gravatt, Thijs Roukens and Wander Wadman.