Inventory forecasting has been computerised for forty years. What has changed recently is not that software does the arithmetic, but what the arithmetic is allowed to look at and how often it runs.
The older approach took monthly sales history for an item, smoothed it, projected the trend forward and applied a safety stock multiplier. It worked reasonably well for steady products and poorly for everything else. The newer approach looks at daily or weekly demand at the level of an individual item in an individual location, learns patterns across similar products rather than treating each one in isolation, and produces a range of likely outcomes instead of a single number.
That shift from one number to a range is the part that matters operationally, and it is worth understanding before evaluating any of the tools that claim it.
What Forecasting a Range Actually Buys You
A traditional forecast says an item will sell 400 units next month. A probabilistic forecast says there is a high likelihood of selling between 310 and 520, with the middle of that range around 400. The first answer is more comfortable. The second one is more useful.
The reason is that safety stock exists to cover the spread, not the average. If you know the spread, you can size the buffer to the service level you actually want on that item rather than applying the same rule to the whole catalogue. Items where a stockout loses a customer get covered generously. Items where a stockout is an inconvenience get covered thinly. The total inventory investment often falls while availability on the products that matter improves.
Done without that distinction, forecasting improvements tend to produce a slightly better single number and no change in how much stock is held, which is why some projects deliver very little.
Where These Models Are Genuinely Better
Three situations stand out, and all of them are common in Canadian distribution.
Seasonality that does not repeat cleanly is the first. A product whose demand depends on weather rather than the calendar is hard to forecast from history alone, because the same week in two different years is not comparable. Models that can take in external signals handle this better than a smoothed average ever will.
Intermittent demand is the second. Many industrial and replacement parts sell in ones and twos with long gaps, and classical methods either overstock them or declare them unforecastable. Learning across similar items rather than from one item's thin history is a real improvement here.
Lead time variability is the third, and it is the one Canadian businesses feel most. Replenishment from an overseas supplier to a west coast port, then inland by rail, then out to a regional facility, accumulates variability at every step. A forecast of demand alone is not enough; the useful output is a forecast of demand over the replenishment lead time, including how uncertain that lead time is. Border processing, corridor closures and port congestion all widen that distribution in ways a supplier's quoted transit time does not capture.
What They Are Still Bad At
The honest list is short but important, because every item on it is a situation where a confident forecast is worse than an uncertain one.
New items have no history, so the model is either guessing or borrowing from products a human decided were comparable. Promotions and price changes break the relationship the model learned, unless promotional periods were recorded well enough to learn from, which in most businesses they were not. Structural shocks fall outside the pattern entirely. A tariff change that moves a product's landed cost, or a competitor exiting a category, is not something a demand model can anticipate from sales data.
The failure mode in all three cases is the same. The system produces a number with apparent precision, nobody questions it because it came from a model, and the error compounds through purchasing decisions for a quarter.
The Data Work Nobody Advertises
Most disappointing forecasting projects fail on inputs rather than algorithms. The prerequisites are unglamorous:
- A clean item master, with one record per product rather than several created over the years by different people
- Consistent units of measure, so that a case, an each and a pallet are not silently mixed in the same demand history
- Demand history rather than shipment history, since what you shipped during a stockout understates what customers wanted
- Recorded lost sales and substitutions, which is the only way a model can learn the difference between low demand and no availability
- Actual receipt dates against purchase order dates, so lead time variability can be measured instead of assumed
- Promotional and price-change periods flagged, so unusual weeks can be treated as unusual
A business with those six things in order will get useful results from fairly simple methods. A business without them will not get useful results from sophisticated ones.
What It Changes Inside the Building
Forecasting is usually discussed as a purchasing problem, but the warehouse feels the output as directly as the buyer does.
Better item-level forecasts allow slotting to be reviewed on evidence, so that products about to enter their season move closer to the pick face before the volume arrives rather than after. Labour planning improves for the same reason, because a forecast that predicts a busy Thursday is worth more than one that predicts a busy month. Inbound dock scheduling becomes smoother when expected receipts are known with some confidence, which reduces the trailer congestion that makes everything else slower.
This is where the warehouse and the forecast meet in practice, and it is part of why businesses increasingly want their 3PL warehousing provider connected to the same inventory data they plan from rather than receiving a purchase order and a hope.
Keeping a Human in the Loop
The governance that makes this work is not complicated. Track whether the model beats a simple benchmark, because a surprising number do not. Watch for bias as well as error, since a forecast that is consistently ten percent low is more damaging than one that is noisily accurate. Record the reason whenever a planner overrides the system, so that the pattern of overrides tells you where the model is weak. And keep a person accountable for the plan rather than the tool, because the tool cannot be asked why.
None of that is an argument against the technology. It is the difference between a forecast that improves decisions and one that automates the same decisions faster.
Where to Start
For most Canadian distributors the first gains come from fixing demand history and lead time records, then applying service-level-based safety stock to the items that actually matter. That work pays off regardless of which software is eventually chosen, and it is a prerequisite for all of them.
18 Wheels stores and moves inventory for businesses across Canada, with warehouse and fulfillment operations integrated into customers' systems so that receipts, stock positions and shipments arrive as data rather than paperwork. We also invest in technology across our operations for the same reason our customers do, which is that the decisions get better when the record is accurate.
If you are planning a forecasting or inventory project and want the warehousing side to support it properly, request a warehousing quote or get in touch and we will talk through what your data would need to look like.
