Bol.com is a retail platform with over 12 million clients within the Netherlands and Belgium. To serve the wants of those clients, it’s important that our crew supplies correct forecasts to empower enterprise choices on a everyday foundation. With the outbreak of the COVID-19 pandemic, on-line procuring behaviour went by means of a significant shift and demand for important and non-essential merchandise swiftly elevated (see cbs).
After months of combating producing any cheap forecasts, we managed to design a characteristic which is ready to describe the dynamic modifications related to the COVID-19 pandemic. The addition of this characteristic permits us to offer dependable forecasts within the brief time period and roll-out situation forecasting for the long run, supporting totally different domains throughout the enterprise. Our strategy is well interpretable and explainable to stakeholders, main to higher knowledge pushed choices.
In our forecasting panorama, we offer various kinds of gross sales forecasting, each on complete degree and likewise on product degree. These forecasts are then used on their very own for various functions throughout the organisation but additionally utilized by our crew as the principle 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 answer which might scale throughout the totally different modelling approaches.
We began with an in depth knowledge evaluation and studying on the knowledge relating to COVID-19 restrictions to include the unfold of the virus. With this data 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. exhibits a illustration of the COVID-19 severity index.
This COVID-19 severity index is much like the publicly accessible COVID-19 stringency index (accessible right here). The principle distinction is that our severity index is tailor-made to the actual dynamics of the affect of the pandemic on our gross sales knowledge, 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 accessible stringency index would counsel. Creating our severity index additionally permits us to extra simply translate the brand new data on restrictions from press-conferences right into a future trying index, which is prime for forecasting. As well as, we’re capable of create higher and decrease bounds for the severity index primarily based on the earlier months knowledge and with this allow situation forecasting.
Because the COVID-19 severity index is only a time-series characteristic, 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 newest pandemic dynamics and consequently additionally elevated their belief in our predictions.
For long run planning, we additionally aided our stakeholders by offering totally different situation forecasts, every one primarily based on a special attainable end result of the pandemic. On this course of, stakeholders might request a situation primarily based on a reference interval for which we might then prolong 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 professional data under consideration to resolve on the long run values for the severity index. The truth that the severity index is well 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 answer to forecasting demand throughout a pandemic has allowed us to offer dependable forecasts to empower enterprise choices within the brief time period, and roll-out situation forecasting for the long run. We might not know but what the way forward for the pandemic appears to be like like, however we at the moment are significantly better ready.
*This work was developed as a joint effort from Staff 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.