The Future of Demand Forecasting

A Primer on the Future of Demand Forecasting

Simon Joiner, ML/AI expert and Director of Product Management at o9 Solutions, recently led a discussion with former and current supply chain leaders from Unilever, PepsiCo, Mars, and Converse on how the use of ML/AI technology is revolutionizing demand forecasting. During the discussion, participants were shared their insights, doubts, and experiences with demand forecasting at their organizations,

Here are the main takeaways from the discussion.

1. Traditional Forecasting Methods Require Heavy Lifting

Traditionally, companies have used hard-coded or “fixed” regression models to forecast sales volumes based on promotion drivers and market insights. While these models require significant effort to build, train, and test, their largest drawback is that they cannot adapt to new information and learn from past experiences. Teams must manually intervene and adjust the model to improve its accuracy. Moreover, forecasting sales requires forming a hypothesis about which promotional drivers will impact demand, which is equally time-consuming and subject to human bias.

Explained a former planning leader at Unilever:

“We manually defined all the promotions drivers to forecast sales volumes. We built a huge regression model. I will call this a hard-coded model because we essentially built and derived it from massive amounts of data and correlation factors. But it couldn’t learn by itself. If we wanted to improve it, we had to intervene manually and adjust it. We also had to manually steer and hypothesize about which drivers might be influencing demand. There was a lot of manual intervention required before it worked.”

2. ML/AI Technology is a Powerful Lever for Better, Faster Forecasting

Companies today have access to a wealth of data. But with vast amounts of data comes a different challenge: building forecasting models that are both efficient and yield high-quality results. Historically, cleansing and processing data from different sources and varying levels of quality has been a painstaking process. ML/AI technology, however, has proven to be a powerful lever to quickly collect and analyze massive amounts of data to determine which variables influence demand for products the most. Implementing these systems is relatively straightforward, as companies need only collect, assess, and sort their existing data to get started. By leveraging these advanced technologies, companies can generate accurate forecasts with less effort than traditional regression models.

Said Simon Joiner, an expert in ML/AI and Director of Product Management at o9 Solutions:

“ML/AI technology allows you to collect enormous amounts of data. But that’s not necessarily the point. You want to obtain the data and have your machine learning capability assess it, tell you if it suits your objectives, and leverage it to shape your forecast. And you can start straight away, even if you don’t have a sophisticated ML/AI capability yet. The first step is to start assessing, collecting, and sorting your data in preparation.”

3. Collaboration is the Key in the Demand Forecasting Process

One participant, a former supply chain executive at PepsiCo, recounted an experience in which he worked with customers whose forecasting systems and algorithms were equally robust. This created a discrepancy in the predicted promotional volumes, as each party had an opinion. The participant emphasized the importance of collaboration and education to bridge this gap. The parties could find common ground by sharing knowledge about each other’s models, resulting in a more accurate forecast. This collaborative effort can lead to a more resilient supply chain better equipped to handle disruptions.

Said the participant:

“I think there are so many different systems and solutions out there, but ultimately, collaboration is the key. Because without facts, someone will give an opinion better than yours.” 

4. Organizations Need to Look Beyond Their Walls for Better Accuracy

In a similar vein, Simon Joiner highlighted the dangers of an “inside-out” approach, where companies solely rely on their data and information. This can lead to biases and suboptimal decisions. Simon stressed the importance of working with suppliers and customers to achieve an “outside-in” approach, where all parties share a common data set and make informed decisions.

Explained Simon:

“That’s the ‘inside-out approach’: companies only looking at their data and point of view. This will lead to biases. When you start collaborating with your suppliers and customers, you’re sharing the same data to help each other, and that’s what we want to be: sustainable, economical, and more accurate.”

5. Organizations Don’t Need ML/AI Technology to Get Started

A supply chain executive at retail fashion and apparel company Convense acknowledged that their company faces the challenge of incorporating external factors into their decision-making process, but they currently lack the ability to handle the vast amount of data involved. Simon offered a solution for companies in similar positions, suggesting that they can start collecting and analyzing data from their existing systems without waiting for budget approvals for cloud-based software. Over time, patterns will emerge, and insights can be gained.

6. Inventory Levels are a Key Promotion Driver 

The former supply chain executive at PepsiCo further emphasized the importance of tapping into one crucial data point: inventory levels shared by customers. By analyzing this data, companies can gain valuable insights into the sell-through rates of their promotions.

“The one data point that everyone should be tapping into is the inventory levels their customers are willing to share. That is one of the most critical reasons I sold a thousand cases during last month’s promotion but only 500 cases this month.”