Forecast accuracy by category: what fashion retailers report in 2026

Fashion retailers in 2026 remain reluctant to publish the demand forecasting metrics that determine whether collections sell through or mark down. A survey of annual reports, investor presentations and supply chain conference proceedings from 2025 through October 2026 reveals that specific mean absolute percentage error figures and forecast bias disclosures are rare, even as executives acknowledge that better demand signals directly affect margin and cash flow.

What does emerge is a widening gap between accuracy expectations for different product types. Staple replenishment items show tighter error bands than seasonal fashion, and the gap matters because fashion represents the highest margin opportunity and the highest inventory risk. According to published retail forecasting literature, mean absolute percentage errors for fashion items have long run in the thirty to fifty percent range, versus far tighter bands for staple replenishment.

The lack of public disclosure does not mean retailers have stopped measuring. Internal dashboards track MAPE, bias, and weighted error at style, size and location level. The choice not to publish reflects competitive sensitivity and the difficulty of explaining why a 35% error rate can coexist with profitable operations when the forecast supports buy discipline rather than precision.

What forecast accuracy do fashion categories typically achieve?

Industry benchmarks published in 2025 and 2026 suggest that apparel MAPE at the SKU level typically falls between 25% and 60%, with the range determined by product lifecycle, trend sensitivity and the availability of comparable sales history. According to a February 2026 analysis, apparel and fashion typical weighted MAPE runs 25% to 40%, and new or seasonal launches can run 50% or higher, because style rotation, trend sensitivity and size or colour complexity all work against accuracy.

A June 2025 retail benchmarking study noted that fashion and apparel retailers often consider 30% to 40% MAPE acceptable at the SKU level, though category level accuracy should be considerably tighter. For comparison, grocery and fast-moving consumer goods retailers typically target MAPE below 10% to 15% for high velocity items, highlighting the structural forecasting disadvantage fashion faces.

The wide tolerance reflects the reality that a forecast for a never before seen seasonal item has neither history nor a proxy for history. Published retail case discussions through 2025 show that organisations with defensible forecasting records share three habits rather than one algorithm: they buy in steps, committing early only to a fraction of the season and holding capacity for in-season reaction; they measure error at buy level and publish it internally, which keeps judgment honest and models calibrated; and they integrate returns data, since a fit driven return spike reads as demand to a naive forecast and poisons the next cycle.

Which retailers have disclosed accuracy improvements?

H&M deployed AI driven demand forecasting and inventory management systems during 2025, using historical sales data, market trends, customer preferences and external variables such as weather patterns and local events to generate detailed forecasts. According to a case study published in early 2026, forecast accuracy improved from 68% to 82% within six months in pilot stores, boosting on-shelf availability by 12 percentage points and increasing online conversion by 3.5%, while staff productivity rose 20% through automation of routine tasks such as replenishment requests.

Adidas reported similar progress in pilot implementations during 2025. A comprehensive case study published in October 2025 documented that forecast accuracy improved from 68% to 82% within six months in pilot stores in North America, with plans to scale to Europe and Asia Pacific by the third quarter. The demand to delivery loop integrated wholesale data, retail point of sale and shopper signals, providing weekly updates to the production network.

Levi Strauss partnered with an AI firm to deploy machine learning models that predict demand at the store level, analysing historical sales, local trends, weather and promotional calendars to optimise inventory distribution and replenishment. The implementation resulted in 90% inventory accuracy across stores and a waste rate of 1%, down from a prior range of 5% to 7%, alongside faster sell through and reduced markdowns, though the company did not disclose specific MAPE figures in public filings.

Inditex, parent of Zara, has long operated a pull system in which point of sale data flows from stores back to headquarters and triggers production and replenishment. According to a 2025 supply chain analysis, Zara maintains a low unsold inventory rate of 10% by accurately forecasting fabric needs and production, and its shipping accuracy rate has reached 98.9% with an error rate of less than 0.5%. The company now leans on AI demand forecasting that reads signals like search behaviour, social sentiment and even local weather to decide what to make and where to send it.

How does accuracy vary by category and lifecycle stage?

Forecast accuracy in fashion retail is not uniform. A deep learning study published in early 2025 found that models achieve their strongest relative improvements in trend sensitive segments such as seasonal accessories and limited edition collections, where cold start challenges and censored demand effects are most pronounced. In contrast, stable categories such as staple apparel exhibit consistent but less dramatic improvements, reflecting the inherently lower variability of consumer demand.

According to demand planning guidance published in December 2025, accuracy should be tracked by multiple views: category, because some categories behave more unpredictably; channel, because online versus store demand often follows different patterns; region, because climate, culture and local events affect demand differently; and lifecycle stage, because new launches and end of life items tend to have lower accuracy.

A machine learning study of intermittent sales in fashion retail, using data through the end of 2025, found that two stage models achieve superior accuracy for products with sparse demand patterns by decomposing the forecasting task into demand occurrence classification and conditional demand magnitude estimation, effectively reducing overestimation during zero demand periods.

What improvement ranges does McKinsey cite for AI forecasting?

McKinsey's State of Fashion 2025 research, published in November 2024, found that AI powered forecasting has the potential to reduce inventory by 5 to 15% and improve stockouts by 15 to 25%. Both figures come from closing the same gap: category level forecasts hide the variance that determines whether a specific style, in a specific size, in a specific store, sells through or sits.

A separate 2025 analysis cited by a demand planning vendor noted that AI powered forecasting can cut fashion inventory by 5% to 15% and stockouts by 15% to 25%, and that one implementation delivered 92% forecast accuracy and 78% fewer stockouts, though the specific client was not identified in public materials.

An academic study published in early 2025 found that AI models can cut retail forecast error from a mean absolute percentage error of 28.76% down to 16.43%, a roughly 43% improvement that flows straight through to fewer stockouts and less dead stock. Moving from a naive baseline to a well built AI system commonly reduces MAPE by 30% to 45%, though the exact gain depends on data quality and demand volatility.

The challenge is that these figures represent potential or pilot results rather than disclosed enterprise wide performance. Fashion executives polled in the annual BoF McKinsey State of Fashion Executive Survey in 2024 and 2025 identified customer product discovery and search as the top ranked use case for generative AI in 2025, ahead of demand forecasting, suggesting that deployment priorities still vary across the industry.

Why do retailers not publish MAPE and bias in annual reports?

The absence of forecast accuracy metrics in public filings reflects several factors. First, competitive sensitivity: a retailer that discloses granular MAPE by category signals to suppliers, competitors and analysts exactly where its demand planning capability is strong or weak. Second, the metrics require context to interpret. A 35% MAPE for a seasonal fashion item may represent good performance if the buy process is staged and the retailer holds capacity for reactive replenishment, but the same number looks poor if presented without explanation.

Third, internal measurement frameworks vary. Some retailers track MAPE at the final buy level, others at the initial plan level, and still others measure forecast value added, which compares the accuracy of a statistical or AI model to the accuracy after human planners apply overrides. A category level MAPE can look acceptable while individual styles perform badly, masking the product level variance that drives actual margin outcomes.

Fourth, stockout censored demand distorts the numbers. A product that sold out on day ten of a thirty day period looks like it only sold what was in stock, not what demand actually was, and forecasting models trained on unadjusted history repeat the understatement. Returns and exchange patterns further complicate the picture, particularly in categories with high fit related returns where available to promise stock shifts differently than in categories with low return rates.

What does the McKinsey State of Fashion 2026 report say about forecasting?

The State of Fashion 2026 report, published by McKinsey and The Business of Fashion in late 2025, notes that with turbulent conditions including volatile input costs, supply chain disruptions and slow growth straining fashion's economic model, AI is shifting from a competitive edge to a business necessity. More than 35% of executives report already deploying automation with generative AI in select functions for routine tasks.

The report highlights that excess stock and inventory mismatches remain significant issues, with overproduction valued at 70 to 140 billion dollars annually. Climate driven demand fluctuations further complicate forecasting. To address these challenges, brands are adopting data driven tools for accurate demand prediction and dynamic purchasing models that minimise waste.

According to the BoF McKinsey State of Fashion Executive Survey conducted in August and September 2025, 46% of respondents said they expect industry conditions to worsen in 2026, up from 39% the year prior, citing trade war uncertainties, unpredictable tariffs and shifts in consumer spending as key risks. Only 25% of fashion leaders believe that industry conditions will improve, up from 20% a year earlier.

How do retailers measure forecast bias alongside error?

Forecast bias shows whether the process repeatedly forecasts too high or too low. A low total error can still hide offsetting over forecasts and under forecasts, which is why best practice tracks both MAPE and bias. According to a fashion demand forecasting guide published in mid 2026, at the post season review, planners should identify where error and bias clustered, which assumptions failed, what data was unavailable at the decision cutoff, and what will change in the next baseline.

Low accuracy with negative bias points to under forecasting and service risk, meaning stockouts, lost sales and expediting costs. Low accuracy with positive bias indicates over forecasting and working capital risk, resulting in excess inventory, markdowns and obsolescence charges. High accuracy with low bias represents the target state, though even that requires interpretation: a forecast that consistently hits within 5% but always comes in 4% high still creates bias that compounds over multiple seasons.

Demand planning key performance indicators published in 2026 recommend tracking forecast accuracy, inventory turnover ratio, fill rate and customer service level together, because forecast accuracy alone does not reveal whether the supply plan translated accurate predictions into available product. A retailer can achieve 85% forecast accuracy yet still experience stockouts if the replenishment cycle, safety stock rules or supplier lead times do not align with the volatility the forecast predicts.

What are the risks and limits of published accuracy claims?

Reported accuracy improvements should be read with care. Pilot results in a controlled subset of stores or styles do not necessarily scale to the full assortment, and accuracy measured over a single season may not hold when trend direction reverses. A forecast model tuned to minimise error during a period of steady growth can produce high bias when demand declines, and vice versa.

Vendor case studies, which provide much of the publicly available detail on fashion forecasting performance, typically highlight best case outcomes and may not disclose the share of SKUs or channels where accuracy did not improve. Similarly, percentage improvements are sensitive to the baseline: lifting accuracy from 60% to 75% represents a larger absolute gain than lifting it from 80% to 95%, even though both constitute a 15 percentage point move.

Academic studies and commercial benchmarks often use different error metrics, making direct comparisons difficult. MAPE, weighted MAPE, mean absolute error, root mean squared error and forecast value added all measure different aspects of performance, and a model that optimises for one may perform poorly on another. Fashion retailers that switch metrics between reporting periods or mix figures from different methodologies risk presenting a misleading picture of actual progress.

External shocks further complicate assessment. Forecast accuracy measured in 2024 and 2025 reflects demand patterns shaped by post-pandemic normalisation, inflation, tariff uncertainty and geopolitical disruption. A model that performed well in that environment may struggle if macroeconomic conditions, trade policy or consumer confidence shift sharply, as several fashion executives noted in conference presentations during 2026.

Frequently asked questions

What should fashion companies do now?

1. Measure error at decision level, not reporting level. Calculate MAPE at the granularity where the buy, allocation or replenishment decision is made, typically style-size-location or style-size-channel. Category level error masks the variance that determines margin.
2. Track forecast bias as well as magnitude. Persistent over-forecasting or under-forecasting signals a systematic problem with the model, the input data or the judgment overlays that planning teams apply. Review bias by product type, season stage and planner.
3. Audit history for stockout distortion. A product that sold out early looks like it only sold what was in stock, not what demand actually was. Adjust historical sales to reflect lost demand before training forecasting models, or segment SKUs by availability pattern.
4. Segment SKUs by forecastability, not just by revenue. High velocity basics with long history and stable demand warrant different forecasting approaches, model parameters and safety stock rules than trend-driven seasonal launches with no comparable past sales.
5. Build an attribute taxonomy for new products. Tag every product by cut, fabric, price tier and category so a new style without history can be forecast against comparable past performance instead of a category average. Test attribute based similarity models before the season starts.
6. Set improvement goals from your own backtests. Industry benchmarks provide context, but the most meaningful target is improvement over the brand's own historical baseline at the category and channel level, measured using a consistent methodology and time horizon.
7. Shorten the reforecast cycle for fast movers. If replenishment decisions run monthly, pilot a weekly reforecast for the highest velocity styles. The lift in accuracy on fast movers is usually visible within one reorder cycle and can fund broader deployment.