Every unit you forecast wrong costs money twice. Over-forecast and you tie up cash in stock that ages, gets marked down, or is written off. Under-forecast and you stock out, lose the sale, and often lose the customer to a competitor who had it. Demand forecasting sits directly on the line between working capital and lost revenue, which is why even modest accuracy gains pay back quickly. This article covers how AI improves forecasting for retailers, what it realistically can and can’t do, and how to deploy it without betting the business on a black box.
What a forecast is actually for
A forecast is not a prediction you admire — it’s an input to a decision: how much to buy, when to reorder, where to position stock, when to mark down. So the right accuracy target depends on the decision. A slow-moving, high-margin item that’s expensive to hold needs tight forecasting; a cheap, fast-moving staple where overstock costs little can tolerate a rougher estimate.
This reframes the whole exercise. Don’t chase a single headline accuracy number across the catalogue. Forecast at the granularity of the decision — usually SKU x location x week — and judge each forecast by the cost of being wrong in that specific context.
Where classical methods run out of road
Traditional approaches — moving averages, exponential smoothing, ARIMA — are transparent, cheap and genuinely good for stable, high-volume products with clear seasonality. Many retailers could improve materially just by applying these properly. Don’t skip them; they’re your baseline.
Classical methods struggle when:
- History is short or sparse — new products, long-tail SKUs with intermittent demand.
- Demand depends on many interacting drivers — price, promotions, weather, competitor activity, web traffic.
- Patterns shift — a step-change in demand that a smoothing model takes months to catch up to.
- The catalogue is large — hand-tuning thousands of individual models doesn’t scale.
This is exactly where machine learning helps.
What AI adds
Models that learn across products
Modern approaches — gradient-boosted trees, and global neural models such as DeepAR or Temporal Fusion Transformers — train one model across the whole catalogue rather than one per SKU. A new or sparse product borrows patterns from similar items, which is precisely the case classical methods handle worst. This is the single biggest practical win for most retailers.
Many drivers at once
ML models ingest external and causal features that smoothing can’t: planned promotions and their depth, price changes, holidays, weather, marketing spend, even on-site search and add-to-cart signals as leading indicators. Forward-looking demand often shows up in browsing behaviour before it shows up in orders — and your search and recommendations telemetry is a rich, underused source of that signal.
Probabilistic forecasts
The most valuable shift is from a single number to a distribution. Knowing demand will be “around 100, but with a long upper tail” lets you set safety stock against a target service level instead of guessing. Inventory decisions are fundamentally about uncertainty, so a forecast that quantifies uncertainty is far more useful than a point estimate — even if the point estimate is slightly better on average.
Connect the forecast to margin
A forecast only creates value when it changes an ordering or pricing decision. The chain looks like this:
- Probabilistic demand forecast per SKU and location.
- Service-level policy — how often you’re willing to stock out, set by margin and strategic importance, not uniformly.
- Reorder points and quantities derived from the forecast distribution and lead times.
- Markdown timing informed by forecast decay, so ageing stock is cleared while it still has value.
The payoff shows up as less dead stock, fewer stockouts, and fewer panic markdowns — the three places forecasting error quietly drains margin. In our experience the markdown lever is the most overlooked: clearing slow stock two weeks earlier at a smaller discount beats a deep clearance at end of season.
Deploying without betting the business
Forecasting touches purchasing, finance and operations, so trust matters as much as accuracy.
- Always keep a baseline. Run a simple model alongside the ML model. If the ML version can’t beat it on a held-out period, it doesn’t ship. This discipline mirrors how we approach any model — see calculating AI ROI for eCommerce.
- Backtest honestly. Evaluate on past periods the model never saw, and measure error in a way that reflects business cost (weighting by margin or volume), not just raw percentage error.
- Keep humans in the loop for big bets. Let planners override the model for major buys, but track override accuracy. Often the model is right more often than instinct — and showing that builds adoption.
- Watch for drift. Demand patterns change; monitor accuracy continuously and retrain on a schedule.
- Don’t let promotions poison history. Flag promotional and stockout periods so the model learns true demand, not constrained sales. Treating a past stockout as low demand is a classic, self-reinforcing error.
A worked example
A retailer plans a promotion on a mid-tier product. A smoothing model, blind to the promo, forecasts normal demand — so they under-order and stock out mid-campaign, wasting the marketing spend and frustrating customers. An ML model with promotion depth as a feature anticipates the uplift, recommends a larger buy, and flags that a related accessory will see pull-through demand. The difference isn’t a cleverer algorithm so much as a model that sees the drivers the business already knows about.
Pitfalls to avoid
- Forecasting at the wrong grain — too aggregated to drive ordering, or so granular the data is noise.
- One accuracy target for everything, ignoring that error costs differ wildly by SKU.
- Point forecasts only, leaving no basis for safety stock.
- Ignoring lead times — a perfect forecast is useless if you can’t act on it inside the reorder window.
- Treating constrained sales as demand, baking stockouts into future forecasts.
Conclusion
Better demand forecasting is rarely about one heroic model. It’s about forecasting at the grain of the decision, capturing the drivers you already understand, expressing uncertainty so you can set sensible safety stock, and proving — every time — that the clever model beats the simple one. Get those right and the gains in working capital and margin follow. If you’d like to assess where forecasting error is costing you most as part of a wider data insights effort, get in touch.