Physical demand forecasting
Understanding future demand at a physical store allows for better inventory management and the ability to avoid holding too much inventory or running out of inventory and losing sales.
Overview
Predict the future demand for a product.
TARGET
Amount of a product sold in a given day.
Challenge
The core data for demand will be the list of all transactions at the stores which will need to be aggregated to determine the demand on a given day. This demand will be influenced by multiple sources and this additional data will need to be cleaned, joined and transformed into valuable ML features before going into model training. First, information about promotions (both for this product and overall) needs to be taken into account. Next the amount of inventory needs to be extracted from the database. Next external data including weather and lists of key events in the region needs to be attached to the demand data. Finally, time series features need to be created for this data over multiple time scales to capture the behavior that will lead to accurate demand forecasts. This pre-modeling prep process can be frustrating and time consuming. We are here to help.
Modeling techniques and libraries
Time series modeling
Simply model the individual demands as a time series.
Package:
- Statsmodel.tsa
- prophet
Machine learning modeling
Model the individual demands using traditional machine learning including time series derived features for demand along with the other features that may influence the future demand.
Package:
- Sklearn
- Tsfresh
- sktime
