Challenging predictions
A significant part of our Sales volume is promotion driven – ‘buy two, get one free’, bundle pack here, discount there. Because of this, an accurate Sales forecast is not easy to do, especially when you have a few hundred thousand products to sell across the world. Forget moving averages, bye-bye statistical models. To improve accuracy, we employ a set of Machine Learning models – feed them with the past events and train them to predict the future accurately. (Sometimes modifying the past a bit if need be, just do not tell the models, please). The key is not to achieve an accuracy level once but to sustain it.
Last couple of years have been full of surprises. COVID and Brexit just to name two. Customers were inventing new promo ideas, competitors were racing to win scarce market shares. The only constant thing is change, so the model is never complete. With Python as our weapon of choice and Azure for compute and persistence, we run customer centric ETLs at scale, train models weekly, predict, predict, predict and feed output to ever-more advanced supply systems – keeping P&G Supply Chain "best in class". Sustained excellence.