CAN WE PREDICT RESTAURANT SALES?
After a previous project uncovered an opportunity to create a brand new business, the client needed to answer a pressing question before investing time and money into launching the venture: is it possible to predict restaurant sales better than a human? Thus, a feasibility sprint was born.
The first 1.5 weeks was data engineering work to create a robust end-to-end pipeline including ingestion, cleaning/normalization, feature engineering, and model training and evaluation. This set up the rest of the project for success through experimentation with different features and model parameters to determine how feasible prediction was. We got to work exploring relatively simple models, which were able to outperform baseline human mental models by 20%-30% depending on the restaurant.
With confidence that we could forecast sales accurately enough to be useful, we explored what the implications of having those predictions for user experience might be. We went back to our restaurants to present our findings and continue research, insight into how to communicate our predictions and their uncertainty, and how to get owners to trust those projections.
The success of the Bizzy project inspired the client to continue to develop the concept as an internal startup that eventually went to market in fall 2020.