Better forecasting does not necessarily produce better decisions.
Halfway through the client pitch, I was relieved when I realized we wouldn’t be getting this contract.
For several weeks, I had worked with an internal champion to develop a proposal that connected demand forecasts to the company’s production, inventory, and allocation decisions. The project had risks, it would require new processes, user education, and changes to how several functions worked together. But it addressed a real business problem and had a credible path to value.
During the meeting it became clear that the broader leadership team had a narrower goal in mind: improve forecast accuracy.
That’s well within our wheelhouse, however, unless they updated their decision processes as we had argued for they’d simply see a KPI go up without a corresponding rise in margin.
They listened politely but it was clear that the forecast accuracy metric was their goal come hell or high water. Their assumption was clear, if we forecast better, we’ll just make better decisions, and everyone will be happy.
This company wasn’t an outlier. Nearly every operational leader talks about forecast accuracy. It seems obvious that if we can predict demand more accurately then we’ll make better plans, serve customers better, control inventory, and all of this will show up on the bottom line.
I get it because for years, I believed the same thing myself.
The reality is more nuanced, but teams frequently approach forecasting as the underpants gnomes of supply chain:
Better forecast → ??? → higher profit
Forecasts are important, but they’re only part of the story. Two different forecasts with very different accuracy scores, may lead to exactly the same operating decision. If so, then what is the additional value of the more accurate forecast?
My first supply chain role was as a planner for a German LDPE site where I was quickly frustrated by our lack of forecasting capabilities. If only I knew what customers were going to order and when, I thought, I could build much better plans.
So that’s what I did.
I fired up R, got years of historical data, and built a series of sophisticated forecasting models and quickly outperformed what we were doing, which was a low bar given we weren’t doing much of anything at the time.
The new forecasts were better by nearly every metric. This seemed like a clear win.
I saw some improvements in my plans, but I still had late orders and stockouts. I still had to negotiate with my run plant engineers regarding costly grade transitions when key customers ordered more than anyone had anticipated. I still had feedstock constraints, volatile pricing, order cancellations, urgent sales, and all of the other disruptions and firefighting that make supply chain planning difficult.
I realized that, although the forecast had improved, the decision process around it did not.
Even if I perfectly predicted the monthly demand for a customer, it didn’t say anything more specific about when that demand would arrive. Perhaps I wouldn’t have their product on hand until the end of the month when they needed it in the first week.
It also didn’t say anything about what happens if a higher priority customer needs something sooner, throwing my schedule out of whack to meet commercial goals. Nor did it tell me whether a costly grade transition was actually worth making or which customers would get cut when we found ourselves oversold or a plant went down.
These are all decision system problems, not forecasting problems.
The obvious objection is fair: lower forecast error at relevant time horizons can allow a company to carry less safety stock while maintaining its service level. Improved forecasts may also change how particular products or customers are served and which are produced as make-to-order vs. make-to-stock vs. make-to-forecast.
However, these benefits are only realized through a supporting decision system: the people, processes, policies, rules, models, and decision rights that convert the new information into a different action.
That was the missing element in my own forecasting effort. I could improve it - and I could even update my plans based on it - but I could not change our product or customer segmentation, production strategy, or related policies. The improvement happened in isolation; better information without a reliable way to act differently because of it.
That distinction is easy to miss because, in simpler situations, new information easily informs the action. Tell me tomorrow’s temperature and probability of rain, and my wardrobe decision is straightforward. The number of alternatives is small, the constraints are minimal, and the cost of being wrong is low.
Industrial supply chains are not like that.
Every production decision answers what gets produced, how much, and when, and depends on:
The forecast informs the decision. The decision system weighs that information against the company’s constraints, alternatives, economics, and acts accordingly.
Accuracy metrics ignore what errors cost
There are two other problems with most forecasting improvement projects that frequently get papered over.
First, traditional metrics aggregate away the details: not all errors are created equal. There is a plethora of weightings and approaches - all with their own trade-offs - that have arisen to address this, but the core issues remain.
Consider a case of two forecasts where we produce accordingly. The first where we over-forecast demand for a commodity grade by 100 tons and the second where we over-forecast demand for a highly specialized grade by 100 tons. In the first case, the cost is likely small. We have extra high-volume low margin material on-hand that will likely be sold next month. It will show up as a blip on the inventory chart over time.
In the second case, we may be holding that product for quite some time if the demand does not materialize. Depending on expiration dates, inventory constraints, and carrying costs, this may be a costly mistake, let alone the likely higher cost of producing the specialty grade and schedule disruption to make it.
Depending on your metrics, these mistakes may result in the same accuracy but they have very different business consequences.
The second issue is based on a flawed assumption, enough data, modeling, and effort can eliminate uncertainty.
We don’t live in a Laplacian universe and there are clear, diminishing returns to more data and modeling. In other words, uncertainty isn’t a modeling failure, it’s a permanent feature to be designed around.
Decision systems should accept this and design with uncertainty in mind rather than hope it can be eliminated all together. That means de-averaging your forecast away from point estimates into ranges, distributions, and scenarios of possible future outcomes. Only then can you develop plans that are robust to the inevitable disruptions of equipment, supplier delays, demand changes, and everything else that pops up in an industrial supply chain.
Before launching another forecast-accuracy initiative, ask a more basic question:
Which decision will become better if this forecast improves?
Then make the connection explicit.
Who owns the decision? What alternatives are available? Which physical and commercial constraints matter? What does overestimating or underestimating demand cost? At what point would a different forecast actually change the chosen action? How will the business respond when reality inevitably differs from the forecast?
If those questions cannot be answered, the organization is not ready to invest in forecast accuracy.
A decision-centered project may still conclude that better forecasting is needed. But it will target the accuracy that has economic value: the right products, customers, locations, and time horizons. It will also connect the forecast to the policies, models, workflows, and authorities required to act on it.
Success should then be measured through business outcomes such as margin, customer service, inventory, working capital, schedule stability, or expediting cost, not solely through a statistical score.
RedSynth and our team of industrial decision scientists are trained to do exactly that, build a unified system to balance this set of tradeoffs and make critical operational decisions.
If a more accurate number wouldn’t change what you do, you have a decision system problem, not a forecasting problem. Run this brief diagnostic on your process to see where the value really is.