The Bay - Developing a Demand Forecasting Suite to Help Reduce Risks and Make Efficient Financial De
Name:
Mahdi Nacer
Anna Zhao
Pujitha Vasagiri
Saniya Khan
Country: Canada
Organization: The Bay
Through a digital-first, purpose-driven lens, The Bay helps Canadians live their best style of life. The Bay operates thebay.com featuring Marketplace, one of the largest premium life and style digital platforms in Canada, with a seamless connection to a network of 85 Hudson's Bay stores. The Bay has established a reputation for quality and style through an unrivaled assortment of products and categories including fashion, home, beauty, food concepts, and more. Follow us on our social media channels: Instagram, Facebook, Twitter, and TikTok.
Awards Categories:
- Most Extraordinary AI Maker(s)
Business Challenge:
Demand forecasting helps reduce risks and make efficient financial decisions that impact profit margins, cash flow, allocation of resources, and overall spending. Our legacy forecasting approach had shown limited success over the last few years and the organization needed to upscale its forecasting capabilities to the industry standards.
As such, an ad hoc Data Science team was tasked with the development of our new Demand forecasting suite. The models needed to accommodate different levels of product granularity (category and subcategory) for a total of more than 30 models for the first iteration. The end state would support more than 130 models with short-term and long-term predictions.
Two main challenges faced by the team faced were:
- How to handle data quality and data preparation at scale.
- How to build a modeling pipeline that can dynamically support the addition of new product categories over time.
Business Solution:
First, we reach out to the client success team to get some support and guidance on using DSS automation capabilities and best practices from the Canadian and Indian Dataiku Data Science Teams. We got good advice on how to use partitions to make our code dynamics.
A mix of partition and creative python recipes helped us achieve our goals. The ability to switch between low code and full code implementation offers great flexibility.
As of now, our flow is composed of 341 datasets, 219 recipes, and six flow zones. Below is a view of the flow created as part of this initiative.
Value Generated:
The first deliverables have been completed recently, and it's too early to be able to quantify the benefits associated with forecasting models. However, it's a big accomplishment and milestone for our organization.
By building in Dataiku, we were able to free up roughly five people across planning, buying, and marketing to work on higher-value work instead of supporting the delivery of the monthly forecast. Moreover, our iteration cycle has decreased by a factor of three since we onboarded the Data Science into our DSS instance. We are able to push a new version of the model in a couple of weeks as opposed to a couple of months when analysts were using their local machine.
Value Brought by Dataiku:
One of the main added values we got from using our DSS instance was the ability to quickly prototype, test, and iterate. The data preparation was a big challenge, and we were able to accelerate and automate most of it. We clearly see the value of having a centralized end2end ML/AI platform as opposed to local development. Also, DSS allowed us to bring all the delivery team into one tool, reducing excel sprawl and ensuring that the information is nearly all in one place which helps support collaboration and reduce errors