Most manufacturers land somewhere around seventy percent forecast accuracy and stay there. Not because nobody’s trying, but because the standard approach — historical averages adjusted by sales judgement — has a natural limit. You can execute it well and still hit the same wall.
What Holds the Ceiling in Place
Traditional forecasting extrapolates from your own sales history. That history contains the answer to what you sold, but not why. It can’t distinguish a genuine demand increase from a competitor’s stockout, or a decline from a pricing change, or a flat month from a delayed customer order that will land next week.
It also treats every product with the same method. A stable high-volume line and an intermittent spare part have completely different demand behaviours, and a single approach applied across both will handle one well and the other badly.
And sales overlay, though well-intentioned, introduces systematic bias. Sales teams forecast optimistically when targets are involved and conservatively when they’re being measured on accuracy. Both biases are consistent enough to be measured and corrected — which almost nobody does.
What a Model Adds
Multiple signals. Beyond your own history: customer order patterns and reorder cycles, pricing and promotional activity, relevant economic indicators for your sector, competitor stock positions where visible, even weather for products with genuine climate sensitivity. Any of these individually is marginal. Combined, they explain variance that history alone can’t.
Method selection per item. Rather than one technique across the catalogue, a model can apply the approach that fits each demand pattern — and identify which of your SKUs are intermittent enough that a point forecast is meaningless and a probability range is the honest output.
Bias correction. If a particular sales region has run twelve percent optimistic for three years, that’s a correctable pattern rather than an unfortunate personality trait.
Realistic Gains
Movement from around seventy to the low-to-mid eighties is a common outcome for manufacturers with reasonable data. That may sound incremental. In inventory terms it isn’t — forecast error drives safety stock, and reducing error reduces the buffer needed to protect the same service level. The working capital effect is usually the largest single benefit.
Some products won’t improve at all. Genuinely lumpy demand — large infrequent orders from a handful of customers — is not statistically forecastable, and no model changes that. For those items the answer is customer collaboration and visibility into their pipeline, not better mathematics. Knowing which of your SKUs fall in this category is itself valuable, because it stops you chasing accuracy that isn’t available.
The Unglamorous Prerequisite
Models learn from history, so your history has to be clean. Sales data distorted by past stockouts is the most common problem — the model learns you sold two hundred units when demand was four hundred and you simply ran out. Flagging historical stockout periods before training is tedious and materially improves results.
ticktick.ai forecasts at item level with method selection per demand pattern and feeds results directly into material and production planning.
