A question for Chief Supply Chain Officers: does your forecasting process account for the difference between predicting and understanding?

If not, you're in great company. At Omnifold we work with some of the smartest planning teams at some of the world's largest companies, and most of them know the distinction well. But when it comes time to actually put forecasting into practice, these teams apply a laser-like focus to increasing forecast accuracy. 

We have a different perspective – generating accurate predictions from your historical data is necessary but not sufficient. Any complete model of the business needs to also understand the business processes that generate the data.

A painful real world example

Let’s take an example from one of our customers, which demonstrates high forecast accuracy but poor forecasting. 

• A manufacturer forecasts demand at the SKU-plant level (how much of each product each plant will need to make). Capacity at every plant — machines, labor, raw materials — is planned off that forecast.

• One period, Plant A hit its capacity limit, so an order headed for Plant A was re-routed to Plant B. 

• This sensible operations call completely broke their forecast:

  • When the order was re-routed, Plant A’s recorded demand fell and Plant B’s rose. Clearly the shift in demand was due to the company’s own plan, but their data now showed a shift in market demand. 

  • As a result, the manufacturer’s next forecast predicted lower demand at Plant A, and higher demand at Plant B – and the errors cascaded from there.

  • Capacity for the next period is planned off the corrupted demand forecast. The demand forecast runs about a month behind the capacity plan, so by the time true demand shows up, the plan built to meet it no longer fits.

To us, this is the entire problem with statistical forecasting in one story. Most models understand only data, and fail to contextualize the decisions that affect the data. Without this understanding, your measured forecast accuracy may be high, but this is an illusion - by “correctly” predicting the shift to Plant B, you are just predicting the result of a mistake caused by your own planning. 

Why the GenAI trend won’t fix it

It's tempting to assume the current wave of AI handles this on its own. Every planning vendor now says it's "AI-powered" or "agentic"...is it possible that you can simply point AI at this problem, and let the AI agent map out the processes that generate your data?

Structurally, this approach misses two basic facts

1. You are facing an optimization problem, not an automation problem. No amount of automation will get close to fixing what’s really behind forecast errors. And any LLM-powered agent is focused on automation. In reality, each constraint and decision ripples through the supply chain in a unique way for each unique business. The complexity of the problem, and the value that can be unlocked, warrants an entirely different architecture (similar to how self-driving cars are running a fundamentally different type of AI vs ChatGPT). With the right approach, each supply chain can be modeled and optimized. 

2. Your business changes, and your model must be re-trained to adapt. As your business evolves (new SKUs, new channels, new strategies), it’s not enough to give a text file of the updates to ChatGPT and ask for a better forecast. Your supply chain model should be continuously learning from and retraining on your business.

Omnifold is built to solve this problem

Omnifold builds a new model architecture for each customer — one specific to your supply chain's structure, constraints, and objectives. Not a language model fine-tuned on your data, not a digital twin. We develop a novel mathematical structure of your operations, built by people who designed foundation models at OpenAI, invented algorithms for autonomous vehicles, and hold PhDs from Stanford and MIT.

This allows Omnifold to take an entirely different approach to the troublesome re-routed order:

• Omnifold optimizes across supply and demand. By solving for demand, procurement, labor, storage, and transportation as one integrated system you will unlock tremendous growth. Omnifold creates a demand plan that anticipates how the decisions that follow from it will become the next period’s constraints, and reshape the next period’s demand. In other words, the plan and its consequences are solved as one problem. 

• Omnifold adapts. When a constraint is reached (e.g. a plant hits a limit), Omnifold can identify creative solutions based on its knowledge of your supply chain. For example, if re-routing is infeasible, it can suggest substitute SKUs that might be close enough to fill the gap.

• Omnifold understands. Our novel approach to combining the best of optimization and reasoning AI techniques creates a powerful tool for a supply chain team. It’s an optimization engine that understands your supply chain, your business and your constraints. And it can reason about the reality of your business and how it changes. Simply tell the system what adjustment or change is needed, and it will take care of the rest. In the situation above, Omnifold understands the re-routing is not a true change in demand, and prevents the cascading errors from entering the next forecasting cycle.

In direct comparisons to leading planning software (same data, same SKUs, same periods), Omnifold has:

• Reduced inventory 41% for a leading container manufacturer 

• Delivered 10% inventory reduction and 2% revenue lift per month for a high-growth electronics manufacturer

• For a multi-billion dollar health & wellness manufacturer, Omnifold took an 8-figure excess inventory + expedited shipping cost, and reduced it by nearly 30%

This is because we’ve seen the same failure across consumer products and retail, industrials and agriculture, pharmaceuticals. The fix is always the same.

Don't predict. Understand.