Today Rooftop submitted our contribution to the Kaggle competition M5 Forecasting - accuracy
The overall purpose of the competition is to see if using machine learning improves the day-to-day forecast – 1 to 28 days ahead.
Our experience: It matters. We pay special attention to the impact of days with zero sales in the model. In particular, we model the sales of zero units sold separately from the sales of multiple units sold.
We hereby avoid the potential systematic downward bias in the prediction of units sold due to the censored nature of the data caused by the zero-unit sales.
Furthermore, we identify a range of relevant macroeconomic indicators that help improving the forecasts.
Undoubtedly, there is value for businesses in starting to apply machine learning to predict sales, especially when its architecture is determined in combination with a knowledge from economic theory and from empirical economic trends.
Our team has spent the time well during COVID-19, working on this competition, and we greatly appreciate that Forskerparken, Odense, Denmark, made it possible for us, by lending us extra office space during this period. Thank you to Forskerparken.