A New Demand Forecasting Approach Signals a Bottom-Line Boost
By improving its algorithm, a consumer health company saw better accuracy, less excess inventory, and fewer disappointed customers.
A large consumer health company was experiencing a pandemic-driven decline in forecast accuracy. A supply-constrained environment had led to stores placing overinflated orders that exceeded need, which, despite manual Band-Aids to remedy the issue, had a long-term negative impact on forecast precision. When ConsumerHealthCo’s* existing methods struggled to meet this frequent challenge, the company enlisted our help to improve the accuracy of its demand signal.
A hybrid team of consultants and Advanced Analytics experts worked with ConsumerHealthCo to reboot its demand-planning process by first deepening its understanding of how the company currently forecasted orders, then pinpointing the barriers to better accuracy.
ConsumerHealthCo regularly needed to estimate what its retail store customers would order five months out, both by SKU and customer combination, amounting to roughly 5,000 individual series to forecast—and its statistical forecast involved manually selecting algorithms for all products based on analyst discretion. In addition to being inefficient, this approach diverted focus from the products that would make the biggest business impact.
Problematic “ballooned” ordering only compounded the problem. By working closely with ConsumerHealthCo’s supply chain team, we learned that to compensate for such orders, the company forecasted based on actual shipments to customers, with demand planners adjusting the shipment history upward. Data governance gaps meant adjustments could happen at any time—even years after the original order. While this manual adjustment produced the forecast numbers the planners desired, it prevented the company from establishing a single, stable source of truth.
We worked together to redesign the company’s data collection and modeling process, harnessing machine learning techniques to automatically flag and adjust input data and select an appropriate algorithm for each data series. This new process automatically classified products based on their demand patterns, applied the most suitable forecasting models based on these patterns, and selected the best model for each series.
To validate the effectiveness of the prototype, ConsumerHealthCo conducted an in-market test with sample SKUs. While demand planners replaced the existing statistical forecast with the new algorithm, the trial allowed them to adjust based on their understanding of the business. This created an opportunity to share feedback about their reasons for adjustment, which were then used to fine-tune the process.
Jointly developing this new forecast process established best practices for ConsumerHealthCo’s demand-planning team, including better governance in data collection and cleaning. What’s more, the company built a new prioritization muscle, adding customer and category intelligence where it added the most value, which included applying different degrees of human vs. automated forecasting processes for SKUs of varying importance.
Done well, demand forecasting helps reduce excess inventory, stock-outs, and lost manufacturing capacity, but a “set it and forget it” approach doesn’t work in today’s turbulent climate. To garner the demand planners’ trust in the new algorithm, our team refined by continuously incorporating their feedback while gradually applying the new algorithm to more SKUs.
Ultimately, ConsumerHealthCo was able to improve forecast accuracy by nearly 8 percentage points, with the potential to increase it further through customer collaboration and continued refinement. Scaling the algorithm is expected to reduce working capital by an estimated 22% and create an additional 2% capacity via refocused production.