The single-model trap
Most forecasting tools ship one model. It might be exponential smoothing. It might be ARIMA. It might be a neural network. Whatever it is, it is one model, applied across the full SKU range, with the assumption that one statistical method can handle every demand pattern a manufacturer has.
This assumption is wrong, and it is expensive.
Why one model fails
A fast-moving consumer good with strong seasonality has a different demand fingerprint than a slow-moving spare part. A promotional item with demand spikes has a different fingerprint than a steady-state commodity. A new product with no history has a different fingerprint than a 10-year-old workhorse.
Applying one model to all of them means you are, by definition, modelling at least two of those three patterns badly. The fast mover gets a seasonal model that misses the spikes. The slow mover gets a trend model that over-forecasts zeros. The new product gets a historical model that has nothing to work with.
The ensemble answer
The answer is not a better single model. The answer is a blend of models, picked per series, with the blend weighted by which model has been most accurate on that series recently.
This is what ensemble forecasting means in practice: you run multiple model families per SKU, you track accuracy per model per SKU, and you weight the ensemble toward the model that has been winning. The blend updates as accuracy shifts.
For a fast mover with strong seasonality, the seasonal model gets weighted higher. For a slow mover with intermittent demand, the intermittent-demand model gets weighted higher. For a new product, the ensemble falls back to a cold-start heuristic until enough history accumulates.
What you actually see
On the planner's screen, this looks like one number per SKU: the ensemble forecast. But behind that number is a leaderboard showing which model won for that SKU, what the confidence band is, and what the alternative forecasts were.
The planner does not need to know which model won. They need to know that the number they are looking at is the best available estimate, and that if it is wrong, the system can show them why.
The accuracy question
People always ask: what forecast accuracy can I expect?
The honest answer is: it depends on your data quality and demand patterns. A blanket number like "95% accuracy" is a marketing claim, not a forecast.
What you should expect from an ensemble is not a single accuracy number. You should expect, per SKU, a MAPE, a bias, and a confidence band. You should expect to see which SKUs the system forecasts well and which it does not. You should expect to know where to apply human judgment and where the model is reliable enough to trust.
That is the real value of an ensemble. Not a higher aggregate accuracy. A more honest per-SKU picture of where the model is strong and where it is weak.
What to do about it
If you are evaluating a forecasting tool, ask one question: "Do you run one model per SKU, or multiple models blended per SKU?"
If the answer is one model, you are buying a tool that will work well for some of your SKUs and badly for the rest. The aggregate accuracy will look fine. The per-SKU reality will be a mix of good and bad, and you will not know which is which.
If the answer is multiple models blended, ask the follow-up: "Can I see, per SKU, which model won and what the confidence band is?" If the answer is yes, you are looking at a tool that treats forecasting as a per-series problem, not an aggregate one.
That is the difference between a forecasting tool and a planning tool. The forecasting tool gives you a number. The planning tool gives you a number, the reasoning, and the confidence. Planners need the second one.
This article is part of prompt/ed, the Tensor Analytics education hub. Written for manufacturing practitioners, not AI enthusiasts.