Well, I work in this area now, and since this is upvoted a bit I'll give my thoughts. And I'll assume you're constraining the term "demand forecasting" to how its often used in business contexts....as well as your your recent posts on issues getting RNN/LSTM to work your time-series data. IMO the best tool for most product/service demand prediction tasks is domain knowledge for good feature engineering and for getting your data to be more stationary. Why? Product/service demand forecasting problems often start with only few explanatory variables as well as those variables not explaining the variance well (more precisely, low mutual information) relative to the number of actual factors going into the demand. Contrast this with areas getting more media such as deep reinforcement learning, where states and actions are fully representable/observed (e.g., AlphaGo).
May-2-2018, 10:37:38 GMT