Sales Forecasting with Machine LearningThis week’s Featured Blog Friday comes from our Reykjavik University student intern, Guðbjörn Einarsson aka Mannsi, who has been working closely with our Data Scientist, Agnes Jóhannsdóttir, to implement Machine Learning technology into our AGR software. As always, if you have any questions or comments regarding this blog post, feel free to comment on this blog post, tweet us @AGRDynamics, or contact us here.


AGR Dynamics is certain Machine Learning will play a big role in the future of our business. If you haven’t already read through the other Machine Learning blog posts on Recommender Systems and Introduction To Machine Learning you really should, as they are great. Another area where Machine Learning can be applied is sales forecasting. Here we would like to briefly explain how that works and go through the pros and cons.



The most common approach is to use a method called Neural Network. Neural Networks are designed to mimic how the human brain operates and learns and is one of the most complex and powerful forms of Machine Learning. Before you give up because this sounds too complicated, we want you to instead forget about the term Neural Network and instead imagine you have a box. Yes, a box. The neural network will live inside this box but, for our sake, this is just a box.

Sales Forecasting with Machine Learning


This is, however, no ordinary box, but a forecasting box. It takes as input what you believe influences the sales of an item and outputs the sales forecast for the item – usually the inputs are the past sales of the item.


How in the world is the Forecast Box able to make a prediction into the future, you ask? Ideally there is an oracle or a genie living inside the box that simply knows the answer but since this is unfortunately the real world, we are going to have to train the box. Training involves giving the box input data, asking it to make a forecast, and then telling the box what the actual correct output should have been. We can do this by having the box forecast past sales of an item where we know the actual sales amount. Given enough training, the Forecasting Box will learn how an items sales behaves and will use that knowledge to predict future sales.



There are two pros to this method. First, you don’t need to know anything about the behaviour of past sales because the Forecast Box will figure that out on its own. Second, you can decide which inputs you want to use. Most common approach is to use sales records as long back as you think is relevant but if you have a feeling that weather temperature might be affect sales or even a certain holiday you can also include these variables as inputs to the box.



There are also two cons to this method. First, this is a black box method for forecasting. You get the forecasting and a confidence level for it but that is it. It will not tell you why it forecasts the way it does or which variables are mostly responsible for the prediction. Second, training time can be substantial at the start. This will depend on amount of data and the hardware available but it is something to keep in mind.



We at AGR Dynamics are excited to start utilising machine learning in our supply chain management software suite to make our sales forecasting and inventory optimisation tools even better. As Sales forecasting with Machine Learning methods can make even more accurate forecasts, we very much look forward to implementing it in our software and give even better service to our customers. For more information about our supply chain management software tools or to receive a demo, feel free to contact us here.