skip to Main Content
Machine Learning Supporting Demand Planners

Machine learning can help demand planners make an even greater impact.

A leading consumer packed goods company relies on statistical models to forecast demand for their product. While baseline statistical forecasts are generally on target, the company also leans heavily on a team of demand planners — a critical part of the supply chain — to make recommendations that might further improve accuracy and avoid overages or supply/demand issues.

Demand planners factor in intangible touchpoints such as market/sales intelligence the statistical models would not have accounted for. From there, they make determinations as to whether a statistical forecast should stand as-is or if they should override it up or down.

In many cases, these human override decisions prove highly effective at bettering forecast accuracy. Yet when a large organization relies on such human overrides to make daily operational and executional decisions, scale problems begin to present.

That’s where First Analytics comes in.

We wondered: could artificial intelligence emulate the nuanced, gut decisions that demand planners make each day, and recommend statistical forecast overrides with the same or similar level of accuracy? How can we mimic the human intuition developed over years of experience?

Training the AI

Before setting out to prove if machine learning could match the positive interventions of demand planners, we were clear that we were not looking for a way to replace these valued individuals.

We simply believed that if we could train the AI to approximate demand planner performance and accuracy, planners would be freed up to focus on more high-impact business situations. The company could enhance its processes and workflows and grow with confidence.

Our Approach

Our first step was to train the AI to look at the audit trails of previous interventional situations:

  • What was the original statistical forecast by location/brand?
  • What was the demand planner’s override decision?
  • Did he or she let the original forecast stand or calibrate the order up or down?
  • And, did the override turn out to be a good or bad business decision?

We taught the AI to identify the attributes of a human override decision and then classify whether it was better or worse than the statistical forecast.

We then trained a second model to tell us the direction and magnitude of the override, should one be needed.  By applying this two-way “forecast value add” analysis, the AI learned from the positive interventions demand planners made, while ignoring the bad ones.

A MEASURABLE IMPACT

In a comparative, holdout evaluation of 900,000 actual overrides, the machine learning algorithms and human planners were virtually tied in terms of which override resulted in the more accurate forecast.

Machine learning overrides demonstrated a 12.5% accuracy improvement when compared to standard statistical forecasts.

What this means:

Results showed that machine learning can, in fact, achieve a comparable level of accuracy as their human counterparts.

The scalability challenge of marrying data-driven forecasts with intuition-based planner insight would be achieved. We helped enable more intelligent operational forecast developed with the speed and reach of a digital organization.

Now, the company could turn the machine learning loose and allow it to take some automatable burdens from demand planners — freeing up valuable time and knowledgeable resources. The machine learning would augment a planner’s role and help maximize accuracy, productivity and profitability.

We helped a large packaged goods company apply machine learning to emulate what demand planners do day-to-day, in order to:

  • Improve product forecast accuracy
  • Avoid overages or shortfalls in supply
  • Enable more effective use of the time of human capital
  • Maximize productivity and profitability of the business

“The machine learning can simply augment the role of our demand planner teams and give them an even greater ability to focus their time in areas that would add value.”

– Data Scientist, Leading Consumer Packages Good Company

First Analytics has been helping companies use analytics to solve their most important and complex business problems for over 10 years. Founded by Tom Davenport (Competing on Analytics) and Michael Thompson, it has been First Analytics’ mission to advance businesses in their analytic maturity, whether just getting started or accelerating on the more expert end of the continuum. One-time analysis or end-to-end applications. Big data or small. On-premise or in the cloud. We help make analytics a vital element of your daily operations and a key to sustainable growth and success.

We are First Analytics

First Analytics has staff across the US and is based in the Research Triangle of North Carolina — an epicenter of analytics. We are headquartered at the Centennial Campus Research Park at North Carolina State University.

Call (919) 521-5519 or email today.

Back To Top