/Forecasting_Intermittent_Time_Series

How To Apply Time Embeddings To A Classification and Quantity Forecast In Tensorflow Embedding time along with categorical and continuous features offsets the errors caused by intermittent time series. This event often occurs in manufacturing or retail businesses that distribute through multiple overlapping distributors or stores and regions with a large parts list. The solution in tensorflow is demonstrated for an international specialty fastener manufacturer with over 2000 part items and 200 distributors in 5 regions. The manufacturer needed pricing support at the quote level by part, distributor and region to project whether or not a quote would become an order and the expected quantity to be sold during a successful order so that it could align with their current demand planning methods. The first session half discusses applications and solutions while the session second half explains the feature development and the non-linear tensorflow model in depth. The annotated open source code in a jupyter notebook is provided for reference.

Primary LanguageJupyter Notebook

Watchers