What is SKU-level forecasting, and when do you need it?

By Elie Dufeu, CTO & Co-Founder, Metreecs. Published 14 September 2026.

Product-level forecasting predicts demand for each individual product and variant, size, color, flavor, or material, rather than for a category as a whole. It matters because two products in the same category can sell at completely different rates, and a category forecast averages that difference away.

Naomi ran replenishment for a 40-store home décor chain for six years before she ever saw a forecast broken out by product. Her team planned from category totals: “ceramic vases” as one number, “woven baskets” as another. The totals were usually close. The individual products underneath them were often wrong in opposite directions, one style overstocked by 200 units while a near-identical one sold out in three weeks. Nobody caught it until a stock count showed both problems sitting in the same warehouse aisle.

Key Takeaways

• Product-level forecasting predicts demand per item and variant, not per category, and catches errors that cancel out in aggregate numbers.
• A widely cited supply chain benchmark puts product-level forecast accuracy at 85%, three months out, as standard in fast-moving consumer goods and life sciences.
• McKinsey research finds that forecast quality at the product level can exceed 90% when advanced, ML-based techniques are applied well.
• Average forecast accuracy sits near 55% for CPG companies and closer to 70% for tech and electronics, a gap driven largely by product complexity and variant count.
• Global inventory distortion (stockouts plus overstock combined) costs retailers an estimated $1.7 trillion a year, according to IHL Group, and most of that traces back to forecasts built at too coarse a level.

What is product-level forecasting?

Product-level forecasting estimates future demand for each distinct product and variant a retailer sells, rather than for the broader category it belongs to. A category forecast might tell a buyer that “outdoor cushions” will sell 3,000 units next quarter. A product-level forecast tells the same buyer that the navy 45cm cushion will sell 420 units and the terracotta 60cm version will sell 90, numbers a category total can never surface on its own.

The distinction matters because retail demand rarely spreads evenly across a product line. Color, size, price point, and even shelf placement each shift how a specific variant sells, and those shifts often cancel each other out at the category level while staying very real at the product level.

See how AI-powered demand planning builds a forecast for every product you sell →

Product-level vs. category-level vs. store-level forecasting

Retailers rarely pick one granularity and stop there. Most blend levels depending on the decision being made: category-level for open-to-buy budgeting, product-level for allocation, and product x location for replenishment. The table below compares the three approaches directly.

Why product-level granularity is harder to forecast accurately

Forecasting an entire category is statistically forgiving. Add enough products together and random noise in individual items tends to average out, producing a smooth, stable total. Forecasting one product at a time removes that cushion.

A single product’s sales history is thinner, more irregular, and more sensitive to one-off events: a single influencer post, a stockout at a competitor, a week of bad weather. A supply chain director writing in Gartner’s Peer Community forum put the standard benchmark for well-run FMCG and life sciences forecasting at 85% accuracy at the product level, three months out, noticeably lower than what most teams see at the category level, because that thinner signal is harder to separate from noise. McKinsey research finds that forecast quality at the product level can climb past 90% when advanced techniques, typically ensemble machine learning models that blend statistical and pattern-based methods, are applied well, which is exactly the gap between a naive product-level average and a properly modeled one.

That gap also shows up across industries. Average forecast accuracy sits closer to 55% for CPG companies and around 70% for tech and electronics, according to Statista benchmark data, largely because CPG assortments carry far more variants per category (flavors, pack sizes, seasonal editions) than a typical electronics line does. More variants per category means more thin, noisy series competing for the same modeling budget.

How to forecast demand at the product level without drowning in noise

The instinct when a product-level forecast looks unstable is to zoom back out to the category. That solves the noise problem by throwing away the information that made product-level forecasting worth doing in the first place. A better sequence keeps the granularity and manages the noise directly.

1. Segment products by data density before modeling. Fast-moving products with 12+ months of clean history can support a full statistical or ML model. Slow movers and new products need a different approach (see step 4), not the same model forced onto thinner data.
2. Clean the history before trusting it. Strip out stockout days (zero sales because there was no stock, not because there was no demand) and promotional spikes, or the model will learn the wrong signal.
3. Add context variables, not just past sales. Price, promotion calendar, and seasonality index each explain variance that raw sales history alone misses, particularly for products with fewer than 12 months on record. Combining internal sales data with external signals closes much of that gap for thinner product-level series.
4. Use analog or attribute-based models for new products. A new product with zero sales history borrows its initial forecast from similar products already selling (same category, price band, launch season), then self-corrects as real sales data comes in.
5. Set a review cadence tied to product velocity, not a blanket schedule. Fast movers benefit from weekly or daily forecast refreshes; slow movers can run on a monthly cycle without losing much accuracy.

Across Metreecs’ work with beauty and home décor retailers, the products that break a forecasting model first are almost always the newest and the slowest, the ones with the least data to learn from. Building a separate, lighter-weight process for that segment protects the accuracy of the model built for the rest of the catalog.

Where product-level forecasting breaks down

Product-level forecasting is not free of tradeoffs, and pretending otherwise sets up buyers to distrust the numbers the first time reality diverges from the forecast. Getting the tradeoffs wrong carries a real cost in lost margin and missed sales, which is why they are worth naming up front rather than discovering them mid-season.

Thin data is the most persistent limitation. A product selling two or three units a week at a single store rarely generates enough history for a statistical model to separate demand from randomness, and no amount of modeling sophistication changes that math. One mistake we repeatedly see is treating every product in a catalog with the same model and the same confidence level, when the honest answer is that forecast confidence should vary by product velocity.

Data infrastructure is the second constraint. Product-level forecasting needs consistent, clean records of sales, price, and promotions at the product level across every channel a retailer sells through. A retailer still consolidating point-of-sale data from three different systems will get more value from fixing that pipeline first than from a more sophisticated model layered on top of messy inputs.

Organizational habits are the third, less obvious constraint. Buyers and planners who have spent years working from category totals often keep reviewing forecasts at that same level even after a product-level model is running underneath. The forecast can be as granular as the data allows, but if the weekly review meeting only ever looks at category summaries, the extra precision never reaches a purchase order or a replenishment decision. Rebuilding the review cadence around product-level detail, not just the model itself, is usually what determines whether the investment pays off.

Product x location forecasting: the next step up in granularity

Product-level forecasting answers how many units of a product will sell, not where those units need to be to actually sell. A product selling well overall can still be stocked out in one store and overstocked in another, the exact pattern Naomi’s team ran into, and product-level forecasting alone will not catch it.

Product x location forecasting extends the same granularity to the store or channel dimension, generating a distinct demand signal for each product in each location. Metreecs builds forecasts at exactly this level as part of its AI-driven inventory optimization approach: the store in a high-traffic urban location and the store in a smaller regional market each get their own forecast, calibrated to their own demand pattern.

Elias, who manages supply planning for a franchise network of 60 quick-service restaurant locations, put it plainly after a season of chasing the wrong replenishment signals: two locations three miles apart sold the same menu items at almost opposite ratios, one running out of a seasonal item weekly while the other carried a growing backlog of the same product. A single product-level number, no matter how accurate on average, could not have flagged that split. Once the forecast was rebuilt at the product x location level, both problems showed up in the same week’s report instead of surfacing months later in a write-off.

FAQ

What is the difference between product-level forecasting and category-level forecasting? Category-level forecasting predicts total demand for a group of related products. Product-level forecasting predicts demand for each individual product and variant within that category. Category totals can look accurate even when the products underneath are simultaneously overstocked and stocked out.

How accurate can product-level forecasting be? A widely cited practitioner benchmark from Gartner’s Peer Community forum puts 85% accuracy, three months out, as standard for well-run FMCG and life sciences forecasting at the product level. McKinsey research shows that figure can exceed 90% when advanced, ML-based techniques are applied well and the underlying data is clean.

Do I need product-level forecasting if I only have a few stores? Store count matters less than product and variant count. A single-location retailer with 2,000 active products and variants faces a similar granularity challenge to a larger multi-store chain with a narrower catalog. The deciding factor is how much demand variation exists between individual products, not how many locations sell them.

How do you forecast a new product with no sales history? New products borrow an initial forecast from similar products already selling, matched on category, price band, and launch timing, and the model updates automatically as real sales data accumulates over the first several weeks.

What data do I need before starting product-level forecasting? At minimum, clean historical sales, pricing, and promotional records at the product level, ideally 12 or more months, plus a way to flag stockout days so the model does not mistake missing stock for missing demand.

Conclusion

Category-level forecasts are easier to build and more stable on paper, but they hide the product-by-product variation that actually drives overstock and stockouts. Moving to product-level forecasting, and eventually to product x location forecasting for multi-store retailers, surfaces the errors a category total buries. The tradeoff is real: thinner data, more noise, and a bigger data infrastructure lift. For most retail catalogs, the accuracy gain is worth managing that tradeoff deliberately rather than avoiding it.

Book a demo to see how product x location forecasting would apply to your own catalog.

Sources

• Gartner Peer Community, supply chain director comment on forecast accuracy benchmarking (85% at the SKU level, 3 months out, standard in FMCG and life sciences): gartner.com/peer-community/post/most-effective-way-to-measure-forecast-accuracy-at-level-share-forecast-accuracy-reporting-executive-audience-total-company
• McKinsey, supply chain and AI-driven forecasting research (SKU-level forecast quality exceeding 90% with advanced techniques; 20-50% error reduction from AI-driven forecasting): mckinsey.com
• Statista, forecast accuracy benchmarks by industry (CPG approximately 55%, tech/electronics approximately 70%), 2024 data
• IHL Group, 2026 Inventory Distortion research ($1.7 trillion global cost, 6.2% of global retail sales, 65.6% attributable to stockouts)