Visual trend forecasting for fashion & consumer products

Fashion and consumer-goods teams must commit to colours, silhouettes and quantities 6-18 months before a product is sold, long before any sales history exists. Visual trend forecasting uses vision-language models to tag product attributes in huge image streams from social media and runways, builds a time series per attribute and forecasts which ones will grow or fade. The result informs collection planning, buying and initial production runs, where overstock and missed trends are most costly.

Problem & value

• Who: Fashion brands, retailers, private labels, and consumer-goods makers (footwear, cosmetics, home textiles, accessories) with long design-to-shelf lead times and large markdown risk.
• Today: Trend agencies (editorial forecasts, trade fairs), buyers' intuition, last season's sell-through, and manual social-media scrolling. SKU-level demand forecasting (PRED-001) needs sales history, so it cannot say anything about attributes that are not yet in the catalogue. Text-trend detection ([LANG-016]) misses the visual dimension (a "cut-out dress" or a specific green is seen, not described).
• Why AI helps: Computer vision converts unstructured images into countable attributes; time-series models then project adoption curves. Vendor Heuritech states that its engine tags 2,000+ attributes from millions of social images across 16 regions and claims forecasts up to 24 months ahead [1]. These accuracy claims are vendor-reported (the vendor page itself quotes both 92% and 91%, so they are not a validated benchmark); treat them as marketing until reproduced on your own back-test.
• Benefit logic (not a measured figure): the value lies in better first-order quantities and fewer markdowns; no independent, peer-reviewed ROI figure was found, so measure it yourself with a back-test against past seasons (see KPIs).

Variants

• Luxury / premium fashion: colour, material and silhouette trends by region and age cohort for seasonal collections; the main commercial segment of Heuritech [1] and WGSN [2].
• Mass-market and private label: fast-fashion and discounters track micro-trends (weeks) and feed ultra-short-lead production.
• Footwear, sportswear, accessories, jewellery: same pipeline with category-specific taxonomies.
• Beauty and cosmetics: shades, finishes and looks from selfie/tutorial video.
• Home, furniture and interior products: style, colour and material trends from interior-design imagery and Pinterest-type boards.
• Food and beverage packaging / consumer products: pack-design and flavour-colour cues from shelf and social imagery.
• Resale and marketplaces: listing images and search data indicate which items gain resale value.
• Private use: not a mainstream use case; consumers are served by recommenders ([REC-012], [GEN-010]) rather than forecasts.

How it works

1. image/video sources
2. rights-cleared ingestion
3. deduplication & person-level filtering
4. attribute tagging (VLM / fashion-CLIP)
5. attribute time series per region & segment
6. forecast model (+ external covariates)
7. merchandising dashboard
8. human buying/design decision
9. feedback with actual sell-through

1. Ingestion: licensed data panels or APIs; scraping only where terms and the EU text-and-data-mining opt-outs allow it. Research APIs such as TikTok's are limited to non-commercial researchers [3], so commercial users rely on vendors or licensed data.
2. Attribute extraction: fashion-tuned contrastive models (FashionCLIP [4], Marqo-FashionSigLIP [5], general SigLIP 2 [6]) provide zero-shot tagging and embeddings; a VLM or object detector crops the garment. The attribute vocabulary is a curated taxonomy, kept stable so that time series stay comparable.
3. Weighting: posts are weighted by audience tier (trendsetters vs. mainstream), region and age segment, because the lead-lag between early adopters and mass adoption is the predictive signal. Academic work on this problem forecasts fashion-element time series per user group from Instagram data (KERN, Ma et al. 2020) [7].
4. Forecasting: classical (ETS, ARIMA), gradient-boosted or pretrained time-series foundation models with covariates (e.g. Chronos-2, which supports covariate-informed zero-shot forecasting [8]); outputs are probabilistic with trend-lifecycle labels.
5. Human-in-the-loop: designers and buyers interpret forecasts, overrule on brand fit, and decide quantities. An LLM may generate readable briefs but should cite the underlying counts.

Data preparation

Two data streams are needed: external images and metadata (social media, runway, marketplaces, typically via licensed vendor feeds) and internal data (catalogue with product attributes, sales, returns, planning calendars). Internally these sit in the ERP or merchandise-planning system, the PIM, the DAM or simply in file shares and Excel season plans.

Typical gaps: inconsistent attribute naming (colour "navy" vs "dark blue"), missing product attributes, images without metadata, no clean link between SKU and sales history, scattered moodboards, and unclear usage rights for scraped images.

Concrete steps: define the attribute taxonomy first (PREP-019) with a glossary of colour and style terms (PREP-015); pull sales and returns from the ERP (PREP-005); tidy season plans from Excel (PREP-011); tag the own catalogue with a vision model and verify a sample with designers to build a test set (PREP-018); clarify licensing for external images (PREP-025). Tools: a PIM or spreadsheet for attributes, a VLM for tagging, pandas for time series. Expect about 30-40 % of total effort here. See GUIDE-003.

Implementation paths

Which path for whom

• smb: A - specialised trend-forecasting vendors and panels deliver licensed image data and ready forecasts, which a small label cannot build itself.
• enterprise: B - the forecast should be combined with own sales, returns and assortment data, so a custom pipeline on vision and forecasting APIs fits better than a standalone product.
• developer: C - open vision-language models and open embeddings on HW2 allow tagging large image sets at low cost and without sending licensed images to third parties.

Path A — Buy (off-the-shelf)

Trend-forecasting platforms: Heuritech (Paris, FR; image-based attribute forecasting, up to 24 months, built on social imagery including Instagram, TikTok, Xiaohongshu [1]; EU vendor), WGSN (London, UK; "data-led forecasting" with a Trend Forecast Database, Catwalks AI Dashboard, TikTok Dashboard and an API; says it serves 6,500+ brands [2]; UK vendor, adequacy decision for personal data currently in place but check). Other vendors reported in the sweep include Trendalytics, Edited and Stylumia (not individually verified in this review). Right choice for almost all brands: the value sits in the proprietary panel and years of labelled history that cannot be rebuilt cheaply. Check contract terms on forecast accuracy, regions covered, export rights and exit.

Path B — Build on cloud APIs

Use a vision-language model API (EU regions via Azure OpenAI/Vertex/Bedrock in EU, or Mistral's vision models) for tagging licensed image sets, plus managed forecasting (Amazon Forecast successors such as Chronos-2 via AutoGluon/SageMaker [8], or Google/Azure time-series services). Suited to a data team that owns a rights-cleared image feed (e.g. own social channels, licensed stock/agency imagery, marketplace listings) and wants a custom attribute taxonomy. Main costs are per-image tagging and data licensing; the main gap is the missing panel of early adopters.

Path C — Self-host (open source)

FashionCLIP (MIT [4]), Marqo-FashionSigLIP (Apache-2.0 repository, check the Hugging Face model card for weights [5]) and SigLIP 2 (Apache 2.0 [6]) for embeddings and zero-shot tagging; open VLMs (e.g. Qwen-VL family) for captioning; Chronos-2 (Apache-2.0 [8]), Prophet, LightGBM for forecasting; a vector DB for similarity clustering. One GPU server (HW2: 24-48 GB) is enough to tag millions of images in batch. Effort is dominated by data acquisition and legal clearance, not compute. Useful on-premises when image data is proprietary (own lookbooks, store cameras).

Path D — Custom training

Worth it for companies with a long image-and-sales archive: fine-tune the embedding model on own catalogue attributes and train a model that maps attribute adoption curves to your own sell-through. FashionCLIP was trained on over 700K image-text pairs from one retailer [4], giving a scale reference. Label noise and non-stationary taste mean models need retraining each season.

State of the art (2026)

• Fashion-specific contrastive models outperform generic CLIP on retrieval: on the Marqo benchmark, FashionSigLIP reaches text-to-image Recall@1 of 0.121 vs. 0.077 for FashionCLIP 2.0 averaged over six datasets [5]. SigLIP 2 (Feb 2025) improves localisation and dense features [6].
• Pretrained time-series foundation models with covariates (Chronos-2, Oct 2025 [8]) make zero-shot attribute forecasting practical for small teams, but benchmarks are generic, not fashion-specific.
• Commercial platforms add TikTok as a source and combine data with expert editors [1][2]; the winner is the best panel, not the best model.
• Limits and failure modes: (1) social-media popularity is not purchasing behaviour; (2) bias of the panel (platform, region, influencer-driven, paid promotion, bots); (3) trend reversals driven by exogenous shocks (weather, news, celebrities) are unpredictable; (4) attribute taxonomy drift breaks time series; (5) accuracy claims are rarely audited; (6) forecasts are about attributes, not about your brand's SKU-level sales, so they must be combined with [PRED-001] and price/fit factors.

Shared resources & combinations

• Building blocks: VLM/fashion-CLIP tagging, image embeddings and vector search (shared with [VIS-022] visual search, [VIS-035] visual brand detection and [REC-012] size/fit), time-series forecasting stack (shared with [PRED-001]), web/API ingestion pipelines, GPU batch inference (HW2).
• Natural combinations: [PRED-001] converts attribute forecasts into SKU demand and replenishment; [LANG-016] adds text-trend signals (search queries, reviews); [GEN-010] and [GEN-001] turn trend insights into design drafts and imagery; [PRED-020] and [VIS-037] feed return and condition data back into attribute success; [VIS-036] pre-tests creatives; [DATA-001] covers shared data platform.

Effort, cost & KPIs

• A: weeks to onboard, subscription per brand/category (price not public); effort mainly change management in design and buying.
• B: 2-4 months for a prototype; costs: image licensing, API tagging, 1-2 data scientists.
• C: 3-6 months; one GPU server, data engineer plus ML engineer; high legal-review effort.
• D: 6-12 months and a multi-season archive.
• KPIs: back-tested forecast error (e.g. rank correlation between forecast attribute growth and realised sell-through for last 4 seasons); share of first-order volume that sold at full price; markdown rate and overstock; sell-out of trend-led items; hit rate of "rising" attributes; time from signal to design brief.

People & change

Involve management (budget, risk of mis-buying), the department leads of design, buying and merchandising as process owners, and IT for data feeds. Legal checks licence terms and platform terms for image data; suppliers and data vendors are partners whose contracts matter. Designers and buyers are the end users and must co-define the attribute taxonomy.

Works council: usually no, since the system analyses products and markets, not employee behaviour (§87(1) no. 6 BetrVG). Inform it under §90(1) no. 3 if planning processes change noticeably. Personal data: social-media images may show people, so the DPO should confirm that only attributes, not persons, are stored; a DPIA is rarely needed if faces are not analysed.

AI literacy: support the team with training on forecast uncertainty and bias towards trends of specific platforms (GUIDE-006). Pilot with one category and one season, comparing the forecast to the buyer's own plan. Typical concern: "creativity is being replaced" - position the output as an input to briefs; designers and buyers keep the final decision. See GUIDE-008 and GUIDE-009.

Risks, regulation & compliance

• EU AI Act: analysing product attributes for market planning is not listed in Art. 5 or Annex III; classified as minimal risk. Keep it away from emotion recognition or biometric categorisation of people in images (Art. 5 and high-risk rules apply): tag garments, not individuals. If AI-generated imagery from GEN use cases is published, Art. 50 transparency applies from 2 August 2026 [9].
• GDPR: images contain people (faces, possibly special-category inferences); use crops of items, discard or blur faces, aggregate before storage, do not store user handles. Document the legitimate interest; run a DPIA if profiling-like data are kept.
• Platform terms and copyright: scraping often violates ToS; the DSM Directive (EU 2019/790) allows text-and-data mining subject to rights-holder opt-outs, and the TikTok Research API excludes commercial users [3]. Buy licensed data.
• Business risks: over-reliance on forecasts with unverified accuracy; herding (all brands buy the same trend); mitigate with back-testing, brand filters, small initial lots plus in-season reorder.

Getting started

1. Define the decision: which category, which lead time, which quantity decision (e.g. colour palette of the next autumn range).
2. Back-test: ask a vendor (Heuritech, WGSN) for historic forecasts on 2-4 past seasons and compare with your realised sales.
3. Build a stable attribute taxonomy and map it to your PIM/product master data.
4. If building: test FashionCLIP/FashionSigLIP zero-shot tagging on 10,000 licensed images and review precision per attribute.
5. Run legal review (platform terms, DSM opt-outs, GDPR) before any scraping; set up a feedback loop after each season.

Sources

1. Heuritech, product page (24-month forecasts, 2,000+ attributes, sources, Paris HQ) - https://www.heuritech.com/ (accessed 2026-10)
2. WGSN, AI + Data forecasting products, 6,500+ brands claim - https://www.wgsn.com/ (accessed 2026-10)
3. TikTok for Developers, Research API eligibility (non-commercial researchers) - https://developers.tiktok.com/ (accessed 2026-10)
4. FashionCLIP repository (MIT; 700K+ Farfetch image-text pairs; Chia et al. 2022, Scientific Reports 12:18958) - https://github.com/patrickjohncyh/fashion-clip (accessed 2026-10)
5. Marqo-FashionCLIP / FashionSigLIP benchmark tables - https://github.com/marqo-ai/marqo-FashionCLIP (accessed 2026-10)
6. SigLIP 2 model card (Apache 2.0; arXiv 2502.14786) - https://huggingface.co/google/siglip2-base-patch16-224 (accessed 2026-10)
7. Ma et al., Knowledge Enhanced Neural Fashion Trend Forecasting (ICMR 2020) - https://arxiv.org/abs/2005.03297 (accessed 2026-10)
8. Amazon Science, Chronos forecasting repository (Chronos-2, Apache-2.0, covariates) - https://github.com/amazon-science/chronos-forecasting (accessed 2026-10)
9. EU AI Act Article 50 (transparency obligations, applies from 2 Aug 2026) - https://artificialintelligenceact.eu/article/50/ (accessed 2026-10)