How AI demand forecasting works for products with no sales history
Attribute similarity methods match new styles to historical data by colour, silhouette and fabric, lifting forecast accuracy by 20 to 40 per cent in 2026.

KEY TAKEAWAYS Summary by the editors
- AI systems now decompose new products into attributes like colour, silhouette and fabric, then match them to historical sales patterns of similar items.
- Retailers report forecast accuracy improvements of 20 to 40 per cent for products with zero sales history compared to traditional buyer intuition.
- Cold start forecasting reduces overstock and markdowns while improving availability of genuinely novel styles.
- Attribute tagging requires discipline: incomplete or inconsistent product data undermines model performance and can amplify bias.
Every fashion season brings hundreds of new styles that have never been sold before. Buyers and merchandisers must commit to production volumes and allocate inventory months in advance, yet they have no sales history to guide them. For decades this cold start problem has been solved by human intuition, comparable sales from last season and rules of thumb. In 2026, attribute similarity methods offer a data driven alternative: artificial intelligence systems that decompose each new product into measurable features such as colour, silhouette, fabric weight and trim details, then match those attributes to the sales performance of historical items that share the same characteristics.
The technique is not new in principle. Retailers have long looked for analogues when forecasting demand for a new leopard print blouse or an oversized trench coat. What has changed is scale and precision. Machine learning models can now process tens of thousands of past styles, identify patterns across dozens of attributes simultaneously, and generate probabilistic demand forecasts in minutes. Multiple retailers and technology providers report accuracy gains of 20 to 40 per cent compared with baseline methods, translating into fewer markdowns, better in stock rates and lower working capital tied up in unsold inventory.
This guide explains how attribute based demand forecasting works, where it delivers the most value, which data foundations are required, and what risks remain unresolved.
What is the cold start problem in fashion forecasting?
The cold start problem arises whenever a retailer must predict demand for a product that has no prior sales data. In fashion this happens every season: a new cut of denim, a fresh colour palette, a silhouette inspired by the latest runway trends. Traditional time series forecasting methods rely on historical sales to project future demand, so they fail completely when no history exists. Buyers instead lean on experience, vendor recommendations, pre season test buys and macro level category trends. These methods are subjective, difficult to scale and prone to bias.
Attribute similarity forecasting reframes the problem. Instead of asking what this exact item sold in the past, the model asks which items with similar attributes sold well, in which contexts and to which customer segments. If a new midi skirt in forest green has never been stocked, the system searches for past midi skirts in dark jewel tones, weights the sales performance of those analogues by the degree of similarity, and produces a forecast distribution. The method still relies on historical data, but it borrows patterns from adjacent products rather than requiring a direct match.
How do attribute similarity models work in practice?
At the core of every attribute similarity model is a product taxonomy: a structured catalogue of features that describe each item. Typical attributes include colour family, silhouette or fit, fabric composition and weight, sleeve length, neckline, embellishments, price band, brand positioning and seasonality. Some retailers add qualitative tags such as trend theme, occasion or style tribe. Each historical product in the catalogue receives a vector of attribute values, and the same tagging process is applied to every new style before it enters the assortment.
Once products are encoded as attribute vectors, the model calculates similarity scores. Techniques vary. Some systems use cosine similarity or Euclidean distance in a high dimensional feature space. Others employ collaborative filtering, treating attributes like user ratings. More sophisticated approaches train neural networks to learn which attribute combinations predict similar demand patterns, even when the surface features differ. Colour matching often relies on perceptual colour spaces such as CIELAB rather than simple hex codes, because two shades that look alike to a human eye should be treated as similar even if their RGB values diverge.
The model then retrieves the top analogues: historical products with the highest similarity scores. It aggregates their sales time series, weighting each analogue by similarity and recency, and applies adjustments for context such as channel mix, regional preferences, promotional calendar and macro trends. The output is a probabilistic demand forecast, typically expressed as a range or distribution rather than a single point estimate. Merchandisers can use this forecast to set initial buy quantities, allocate inventory across stores or distribution centres, and plan replenishment triggers.
- 2015-2018Early adopters experiment with machine learning on historical sales time series for core replenishment items
- 2019-2020Cloud platforms integrate image recognition to auto tag product attributes from photographs
- 2021-2022Attribute similarity methods gain traction for cold start forecasting during pandemic supply volatility
- 2023Generative AI tools emerge to synthesise product descriptions and extract attributes from unstructured text
- 2024-2025Retailers pilot real time demand sensing that updates forecasts daily using search, browse and social signals
- 2026Attribute based forecasting becomes standard practice for new product introduction at large multi brand retailers
Source: Industry reports and company disclosures
Which attributes matter most for forecast accuracy?
Not all attributes carry equal predictive weight. Empirical studies and practitioner experience suggest that colour, silhouette and fabric are consistently among the strongest signals, but their relative importance varies by category and customer segment. For womenswear tops, neckline and sleeve length often matter more than fabric weight. For denim, fit and wash are paramount. For outerwear, insulation type and length dominate. Price tier and brand positioning also play a role: a luxury cashmere sweater and a fast fashion acrylic pullover may share the same colour and silhouette, yet their demand drivers differ.
Colour matching deserves particular attention. Perceptual colour similarity does not always align with sales similarity. A navy blazer may sell like a black blazer in corporate workwear, but a cobalt blue evening dress behaves differently from a midnight blue gown. Context and category modulate the effect. Some retailers solve this by clustering colours into families such as neutrals, brights, pastels and prints, then fine tuning within each cluster. Others use image embeddings from convolutional neural networks, which learn colour representations that correlate with purchase behaviour rather than relying on predefined palettes.
Seasonality and trend timing add another layer. A puffer jacket in millennial pink launched in autumn 2023 is not a perfect analogue for a puffer in sage green launched in autumn 2026, even if the construction is identical, because the broader trend cycle has shifted. Advanced models incorporate temporal decay functions that down weight older analogues, or they segment training data by fashion season to ensure that only recent patterns inform the forecast.
What accuracy improvements do retailers actually see?
Published case studies and vendor claims cite forecast accuracy gains in the range of 20 to 40 per cent for cold start products, measured as reductions in mean absolute percentage error or increases in forecast value added compared with naive baselines such as category averages or buyer estimates. Translating accuracy into business impact depends on the retailer's operating model, margin structure and supply chain flexibility.
One concrete example comes from a 2025 industry presentation. A European multi brand fashion retailer implemented an attribute similarity forecasting system for new seasonal styles across womenswear and menswear. The company reported that forecast error for products with no sales history fell by approximately 30 per cent in the first full season, leading to a measurable reduction in end of season markdown rates and an improvement in full price sell through. The retailer attributed the gain to better alignment between initial buy quantities and actual customer demand, particularly for mid tier styles that were neither obvious bestsellers nor niche experiments.
It is important to distinguish between forecast accuracy and commercial outcomes. A model can be statistically accurate yet fail to deliver profit if the organisation does not act on the forecast, if supply lead times are too long to react, or if the product assortment itself is poorly conceived. Conversely, even a modest accuracy improvement can generate significant value in high volume, low margin categories where inventory holding costs and markdown exposure are large.
| Metric | Traditional buyer forecast | Attribute similarity model | Improvement |
|---|---|---|---|
| Mean absolute percentage error | 45-60% | 30-40% | 20-40% reduction |
| Forecast bias (over or under) | Often positive (overbuy) | Closer to zero | Balanced inventory |
| Time to generate forecast | Days to weeks | Minutes to hours | 10-100x faster |
| Scalability (SKUs per season) | Hundreds (manual limit) | Tens of thousands | High volume feasible |
What data infrastructure is required?
Attribute similarity forecasting demands clean, consistent and comprehensive product data. Every SKU must be tagged with the same taxonomy, using controlled vocabularies rather than free text. Colour names must map to a standard palette. Fit descriptors such as slim, regular and relaxed must be applied uniformly across buyers and seasons. Fabric compositions must be recorded in a machine readable format, not buried in unstructured vendor descriptions. Any gap or inconsistency in the attribute catalogue degrades the model's ability to find valid analogues.
Many retailers discover that their legacy product information management systems were designed for catalogue publishing and regulatory compliance, not for machine learning. Attributes are missing, duplicated across fields, or entered inconsistently by different teams. Remediation requires both technology investment and process change. Some organisations deploy computer vision models to auto tag attributes from product photographs, or natural language processing to extract features from vendor spec sheets. Others mandate stricter data governance, training buyers and merchandisers to complete attribute forms at the point of product creation.
Historical sales data must also be reliable and sufficiently granular. The model needs to know not just that a style sold 5,000 units last season, but how demand varied by week, by store cluster, by channel and by promotional event. Stockout periods must be flagged, because observed sales are censored by inventory availability. Returns and cancellations should be tracked separately from gross sales. Ideally, the retailer also captures near miss signals such as search queries for out of stock items, waitlist sign ups and abandoned baskets, which reveal latent demand that never converted into a transaction.
How do companies integrate AI forecasts into the buying workflow?
Technology alone does not change outcomes. The forecast must be embedded in the decision making process, with clear accountability and feedback loops. Leading retailers treat the AI forecast as one input among several, not an oracle. Buyers review the model's output alongside their own market knowledge, vendor relationships and seasonal strategy. When the model and the human disagree, the organisation must have a protocol: does the buyer have the authority to override, or does the model set a floor and ceiling within which the buyer can manoeuvre?
Some companies adopt a test and learn approach. They allow the model to set buy quantities for a subset of new styles, chosen at random or in categories where human performance is weakest, while buyers retain control over the rest of the assortment. After the season, the team compares sell through rates, margin and stock health between model driven and human driven buys. This experimental design builds trust and surfaces cases where the model excels or fails.
Feedback is critical. When a forecast misses badly, the team should investigate why. Was the analogue selection flawed? Did an external shock such as weather or a viral social media moment shift demand? Was the product photography misleading, causing the tagging algorithm to misclassify the style? Each post mortem feeds into model refinement and data quality improvements. Over time, the system learns from its mistakes, and the organisation learns which contexts favour algorithmic versus human judgment.
What are the risks and limitations?
Attribute similarity forecasting works best for evolutionary products: new styles that remix familiar elements in fresh combinations. It struggles with genuinely disruptive innovation. If a retailer launches a category that has no historical analogue, such as a wholly new garment type or a collaboration with a celebrity whose aesthetic differs from the brand's archive, the model has little to anchor on. In these cases, human intuition, test markets and pre order campaigns remain the primary tools.
Bias is another concern. If the historical assortment skewed toward certain demographics, price points or aesthetics, the model will replicate and potentially amplify that bias. A system trained on years of conservative colour palettes may under forecast demand for bold, experimental hues, discouraging buyers from taking creative risks. This creates a feedback loop: the model recommends cautious buys, the retailer stocks less adventurous product, sales data for adventurous product remains sparse, and the model continues to favour safe choices. Breaking the cycle requires deliberate intervention, such as reserving a portion of the assortment for high variance experiments that are excluded from automated forecasting.
Data privacy and competitive dynamics also matter. Retailers that share attribute and sales data with third party platforms gain access to benchmarking and cross retailer insights, but they risk leaking proprietary information about product strategy and customer preferences. Vendors and technology providers must offer robust data governance, ensuring that one retailer's data never directly informs another's forecast. Regulatory scrutiny of algorithmic decision making is increasing, particularly in the European Union, where transparency and explainability requirements may compel retailers to document how forecasts are generated and how they influence inventory and pricing decisions.
| Data availability | Product novelty | Category stability | |
|---|---|---|---|
| Core replenishment items | high | low | high |
| Seasonal fashion with familiar silhouettes | high | medium | medium |
| Trend led fast fashion | medium | medium | low |
| New category or brand collaboration | low | high | low |
| Limited edition capsule collections | low | high | low |
0 = poor fit, 3 = excellent fit
Source: Author synthesis of practitioner guidance
Frequently asked questions
What should fashion companies do now?
- Audit your product data taxonomy. Conduct a systematic review of attribute completeness, consistency and coverage across all SKUs. Identify gaps and standardise naming conventions before attempting to train any model.
- Start with one category. Pilot attribute similarity forecasting in a single product category where you have rich historical data, frequent new introductions and measurable pain points such as high markdown rates or stockouts. Prove value before scaling.
- Build cross functional ownership. Assign joint accountability to merchandising, data science and product information teams. Forecasting accuracy depends on collaboration, not on handing the problem to IT and waiting for a dashboard.
- Invest in image and text AI. Deploy computer vision and natural language processing to auto tag new styles as they are created, reducing manual data entry burden and improving tagging consistency. Validate outputs with human review initially.
- Establish model governance. Define clear rules for when buyers can override forecasts, how discrepancies are escalated, and how model performance is measured and reported. Ensure transparency and accountability at every step.
- Reserve budget for experimentation. Protect a portion of your assortment from algorithmic forecasting to test bold, high variance styles that may not fit historical patterns. Use these experiments to discover new demand signals and refresh your training data.
- Monitor for bias and drift. Regularly audit forecast outputs for demographic skew, price tier imbalance and category stagnation. Update training data, re weight analogues and adjust similarity metrics to counteract bias and adapt to evolving customer preferences.
Frequently asked questions
Can attribute similarity forecasting replace human buyers?
No. The models provide demand forecasts, but buyers still make strategic decisions about assortment composition, vendor negotiation, brand positioning and creative direction. The technology augments human judgment rather than replacing it, particularly for genuinely novel or disruptive styles.
How long does it take to implement an attribute based forecasting system?
Implementation timelines vary widely depending on data maturity. Retailers with clean product taxonomies and robust historical sales databases can deploy pilots within three to six months. Those requiring extensive data remediation may need 12 to 18 months before achieving reliable forecasts at scale.
What is the minimum amount of historical data needed?
Most practitioners recommend at least two to three years of sales history covering multiple seasonal cycles and a few thousand distinct SKUs per category. Sparse data sets yield unreliable analogues and high forecast variance, limiting the practical value of the model.
Sources
- McKinsey & Company: The state of fashion technology 2025
- Vogue Business: AI and machine learning in fashion forecasting 2025
- Reuters: Fashion retailers adopt AI for inventory management 2025
- Retail Dive: Demand forecasting technology trends 2026
- FashionUnited: How AI is solving the cold start problem in fashion 2025
