The client is one of the largest US companies that caters meals to kids in the schools. They work with thousands of schools and supply various food items on a daily basis. Their procurement has to order just enough ingredients as almost all the items they procure are perishable.
The ask was the ability to predict, with a certain degree of accuracy, the demand for each product 15 months in advance for appropriate hedging. They were using a semi-manual, intuitive estimate and wanted to replace it with an ML engine.
Data auditing: Being a traditional brick and mortar, the company did not maintain databases appropriately. We needed to discuss with IT teams, business users to guess and redo the names of attributes, relationships, and the corresponding ER diagrams. In the later stages of modeling, we found that they were not collecting certain important attributes and manually transforming certain other attributes. INSOFE created collection pipelines and automated the transformation process saving a substantial amount of time.
Modeling: We initially, built a global model (time series) for predictions. However, as there are several nuances at the school level, this was not accurate enough.
To have INSOFE faculty and data scientists solve your business problems, prep your engineering teams to face the real world complexities, visit here
We built a predictive model for each item at the school level and added them up to predict the overall demand. However, the time-series data for each item was not enough at each school level. We engineered several causal features. Some of them were external (weather) and we collected them from relevant sources. We averaged global and local models.
The model performance still was not good. We realized that the local models were erring a lot when there were no orders for a particular item.