AI demand forecasting: past the single reorder number.

The problem

A blanket seasonal multiplier (like adding 30% in November) gets applied to the whole catalog, even though a scarf and a phone case don't share the same holiday curve, so some SKUs stock out in peak season while others sit overstocked.

How it's done manually

Someone eyeballs last year's sales chart, applies a rough seasonal bump to this year's baseline forecast, and adjusts it again by feel if the season looks unusual.

The AI solution

A model learns each SKU's own historical seasonal curve, including which weeks it actually accelerates and by how much, and applies that specific pattern instead of a single catalog-wide assumption.

Example workflow

A holiday-driven SKU's forecast automatically scales up starting the week its own history shows demand accelerating in past years, not on a fixed calendar date applied to every product.

Business impact

Peak-season stockouts drop on the SKUs that actually swing seasonally, while stable SKUs stop getting an unnecessary seasonal bump added to their forecast.

Estimated ROI

The payoff concentrates on categories with real seasonal spread; a catalog that sells evenly year-round sees comparatively little benefit from this specific piece.

Common mistakes

Applying one seasonality curve to the entire catalog, or worse, to a brand-new SKU that has no seasonal history of its own to learn from yet.

Best practices

Group SKUs into seasonality tiers first (evergreen, moderately seasonal, sharply seasonal) and only build a per-SKU curve for tiers where the pattern is strong enough to be worth modeling.

The problem

A planned sale or promotion creates a demand spike that a normal forecast doesn't account for, so teams either under-stock for the sale or over-order based on a rough guess of how big the lift will be.

How it's done manually

Someone estimates the promotional bump from memory or a rough rule of thumb, like assuming sales usually double during a 20%-off event, without checking whether that held true for this specific SKU last time.

The AI solution

A model estimates the expected lift for a specific planned promotion by looking at how comparable past promotions performed on that SKU or a close analog, before the sale runs.

Example workflow

A 20%-off weekend is scheduled two weeks out; the model pulls the lift size from the last three comparable promotions on that SKU and outputs a pre-sale stock recommendation instead of a guess.

Business impact

Fewer promotional stockouts on the SKUs actually being promoted, and less leftover stock from an oversized order built on a rough guess of demand.

Estimated ROI

Brands running a regular promotional calendar see the clearest benefit, since the model has comparable past events to learn from; a brand's first-ever promotion has to lean more on analog data.

Common mistakes

Using one flat lift multiplier for every promotion type, when a site-wide sale, a flash sale, and a bundle deal usually produce very different lift patterns on the same SKU.

Best practices

Tag past promotions by type (flash sale, site-wide, bundle, influencer-driven) so the model can match a new promotion to the right category of past lift, not just any past sale.

The problem

A brand-new SKU has no sales history at all, so the first order quantity is usually a guess, and getting it wrong either strands cash in unsold new stock or causes a stockout on a product that's just starting to gain traction.

How it's done manually

Someone picks a launch quantity based on gut feel, what a similar product sold in its first month, or simply how much budget is available to spend on the first order.

The AI solution

A cold-start model borrows demand patterns from analog SKUs (similar category, price point, or attributes) and blends in any pre-launch signal, like waitlist size or early ad engagement, to produce a starting forecast instead of a guess.

Example workflow

A new SKU launches in a category with three established analogs; the model blends their early-life demand curves with the new product's pre-launch signal to recommend an initial order quantity and a reorder timeline for the following weeks.

Business impact

Launch quantities land closer to actual early demand, so new products are less likely to stock out right as they gain traction or sit overstocked from an oversized first order.

Estimated ROI

Brands launching SKUs frequently see the most value, since the model improves as more launches feed it comparison data; a brand launching once a year has less analog history to draw from.

Common mistakes

Picking analog SKUs based only on category, ignoring price point or audience differences that actually drive very different early demand curves.

Best practices

Choose analogs on multiple dimensions (category, price, target customer), not just the product type, and revisit the forecast quickly once the first few weeks of real sales data come in.

The problem

A single company-wide forecast gets split evenly, or by rough historical share, across warehouses, which ignores that demand for the same SKU can differ meaningfully by region or fulfillment node.

How it's done manually

Someone takes the total forecast and divides it by a fixed percentage per warehouse based on rough historical shipping volume, rarely revisited as regional demand shifts.

The AI solution

The forecasting model runs at the location level, learning each warehouse's own demand pattern per SKU instead of splitting one total number by a fixed ratio.

Example workflow

A SKU's East Coast warehouse shows accelerating demand while the West Coast node stays flat; the model adjusts each location's forecast independently instead of applying the same growth rate to both.

Business impact

Stock lands closer to where demand actually is, which reduces both regional stockouts and the internal transfers needed to fix a misallocation after the fact.

Estimated ROI

The benefit scales with the number of warehouses and the regional variation in demand; a single-warehouse brand doesn't need this piece at all.

Common mistakes

Splitting the forecast by a fixed historical percentage and never revisiting it as new warehouses, channels, or regional demand shifts change the picture.

Best practices

Recompute each location's share on a regular cadence rather than locking in a fixed split, and treat a new warehouse's first few months as its own cold-start problem.

The problem

A forecasting model that was accurate at launch quietly drifts as the business changes, new channels get added, and demand patterns shift, but nobody notices until orders start looking obviously wrong.

How it's done manually

Forecast accuracy gets checked informally, if at all, usually only after someone notices a stockout or overstock that the forecast should have caught.

The AI solution

An accuracy-tracking layer compares each forecast against actual sales after the fact, flags SKUs where the model is consistently off, and triggers a retrain or a manual review before the error compounds.

Example workflow

A SKU's forecast has under-shot actual demand for three consecutive weeks; the system flags it for review instead of letting the same biased forecast keep driving reorder decisions.

Business impact

Forecast quality gets caught and corrected while it's a minor miss, not months later when it's shown up as a stockout or a pile of overstock.

Estimated ROI

This piece pays for itself mostly in avoided compounding error; a small bias caught early costs far less to fix than the same bias left running for two quarters.

Common mistakes

Building a forecasting model once and never checking its accuracy against actual results, treating it as a finished project instead of a system that needs monitoring.

Best practices

Track forecast error per SKU on a recurring basis, not just in aggregate, since a good average can hide a badly wrong forecast on a specific important SKU.