3 facts about time series forecasting that surprise experienced machine learning practitioners.

Time series forecasting is something of a dark horse in the field of data science: It is one of the most applied data science techniques in business, used extensively in finance, in supply chain management and in production and inventory planning, and it has a well established theoretical grounding in statistics and dynamic systems theory. Yet it retains something of an outsider status compared to more recent and popular machine learning topics such as image recognition and natural language processing, and it gets little or no treatment at all in introductory courses to data science and machine learning. My original training is in neural networks and other machine learning methods, but I gravitated towards time series methods after my career led me to the role of demand forecasting specialist. In recent weeks, as part of my team's effort to expand beyond traditional time series forecasting capabilities and into a borader ML based approach to our business, I found myself having several discussions with experienced ML engineers, who were very good at ML in general, but didn't have much experience with times series methods. I realized from those discussions that there were several things specific to time series forecasting that the forecasting community takes for granted but are very surprising to other ML practioners and data scientists, especially when compared to the way standard ML problems are approached.