How to pair WGSN-style direction with real-time AI signals for buys and allocation.
Traditional services such as WGSN remain invaluable for aesthetic direction, macro context, and cross-category narrative. But buys fail when insight doesn’t translate into the right assortment and size mix, timed to how taste moves across regions and channels. An AI-augmented approach complements editorial perspective with real-time social and runway signals, corrected sales, returns reasons, and availability. The goal is practical: fewer stockouts in core sizes, fewer stranded tails, and higher full-price sell-through. Start with signals merchandisers trust. Correct historical sales for stockouts so empty shelves don’t masquerade as “low demand.” Ingest returns reasons with detail (“too tight in thigh,” “gapes at bust,” “heel slips”) to detect misfit by block/last that distorts demand. Add regional trend momentum from visual/search signals, and segment by channel (DTC, wholesale, marketplace) because curves differ by surface. Providers explain how runway/social data predicts timing; see Heuritech. Keep a fashion attribute spine—silhouettes, rises/lengths, fabrics and stretch %, palette, toe/heel shapes—so outputs talk like buyers, not black boxes. Context matters in 2025. Demand is uneven and value-seeking behavior is rising, making precision in buys a margin lever. Leadership teams keep citing the BoF x McKinsey State of Fashion work to frame this shift; review the overview at Business of Fashion and the full collection at McKinsey. The upshot: merch teams need continuously refreshed, explainable signals that map to buy and allocation decisions—not just quarterly decks.
A credible system has three layers: better signals, an explainable model, and a cockpit buyers can operate. Signals: 1) availability-aware sell-through; 2) returns reasons with fit/style context; 3) regional trend momentum; 4) seasonality; 5) channel mix; and 6) intent (pre‑orders, waitlists, wishlists). Model: estimate censored demand to correct for stockouts; borrow learning from sibling styles (same block/fabric/brand) to handle cold starts; localize curves by region/channel; and incorporate intent signals as demand with confidence bands. Output: pre‑season buy ranges by style/region/channel with uncertainty, plus in‑season weekly refreshes that trigger reallocation where it prevents markdowns. Glue the loop with a fashion-grade vocabulary. Map trend attributes to your product spine—silhouettes, rises/lengths, palette, toe/heel shapes—so a recommendation can carry a reason code: “espresso palette rising in EU; wide‑leg peaking; allocate sizes 2–6 heavier.” Treat photography as data (finish, drape, strap width, hardware tone) so edits, search, and “shop the look” align to what buyers are actually planning. Finally, pair forecasting with size/fit intelligence. When buy plans reflect likely size curves and fit risk, PDPs can emit a single, confident size recommendation with a short rationale—curbing bracketing and reducing returns. Evidence-based PDP patterns are summarized by Baymard, and they pair well with allocation tuned by forecast outputs.
Run the program like a P&L. Outcome KPIs: stockout reduction on top sizes; markdown rate/depth on tail sizes; full‑price sell‑through; GMROI; weeks of supply; and transfer efficiency (% reallocated pre‑markdown). Journey KPIs: time from signal to action (waitlist spike → buy or allocation change), exposure lift for size‑rich variants, and conversion deltas where localized edits appear. Attribute at the node, not the channel. Experiment design: start with one category (dresses, denim, or sneakers) in two regions. Prefer randomized stores/sites or matched cohorts with pre‑registered stop‑loss thresholds. Publish weekly readouts with reason codes so buyers can accept/override with confidence. Seasonal playbooks: pre‑season, convert slopes into buy ranges with uncertainty bands; in‑season, refresh weekly and coordinate creative so size‑rich variants get visibility; post‑season, feed back which attributes predicted stalls and which accelerated sell‑through. Reliability and governance make it durable. Set weekly refresh SLAs (faster for fast fashion), keep an immutable decision log (inputs, reason codes, outcomes), and version models so forecast drift is visible. Maintain privacy by design—evaluate consent at activation for any personalized block and minimize PII in payloads. With explainable signals mapped to a buyer-friendly cockpit, AI trend forecasting goes from inspiration to inventory—and the margin follows.