# Forecasting trends with AI: from runway signals to demand curves
In 2018, the Paris-based startup Heuritech began scanning millions of public Instagram and street-style photos, tagging garments piece by piece: sleeve length, silhouette, color, print. When oversized blazers started appearing on influencers and runways, the system flagged a rising signal months before the trend hit mainstream retail. Buyers who acted early stocked the right quantities. Buyers who waited chased inventory or missed the window entirely.
That is the promise of AI trend forecasting: turning fuzzy visual and social buzz into a number you can actually buy against.
In fashion, the "buy" is the commitment. Months before a season, merchandisers decide how many units of each style, color, and size to order. Get it right and you sell at full price. Get it wrong and you either stock out (lost sales) or drown in markdowns (lost margin).
Traditional forecasting leaned on last year's sales plus a merchant's gut. That works for basics. It fails for trends, which move fast and start outside your own sales data, on social feeds, red carpets, and runways.
AI closes that gap by reading external signals early.
Trend forecasting AI stitches together three kinds of input.
Computer vision is AI that "sees" images and identifies objects and attributes. Trained on labeled fashion photos, it can look at a runway image and output tags: "double-breasted blazer, oversized fit, earth tone, wool texture."
Heuritech's core product does exactly this across social media and runway imagery, producing a growth curve for each attribute. The output is not "blazers are trending." It is "oversized blazers in neutral tones are growing X percent among a specific audience segment, projected over the next 12 months."
Google Trends is a free tool showing relative search interest over time. Search "oversized blazer" and you see whether curiosity is rising, seasonal, or fading. It is a blunt instrument, but it is public, free, and a useful cross-check on a paid vision tool.
Layer in social engagement: hashtag volume, saves, and shares. A garment people save is a garment people intend to buy.
External buzz tells you what is rising. Your internal data tells you how buzz converts to sales for *your* customer. A luxury brand and a fast-fashion brand can see the same runway signal and translate it into very different buy quantities.
The magic is in combining all three. External signals give direction. Internal data gives magnitude.
Here is the mental model. You are converting a rising attribute into a quantity.
Think of it in stages:
1. Detect: Vision AI flags "oversized blazer, neutral" rising among your target segment.
2. Quantify the signal: Convert the growth rate into a projected demand index (say, 130 versus a baseline of 100).
3. Anchor to your data: Find comparable past styles. How did a similar silhouette sell for you last year?
4. Scale: Apply the signal index to your baseline to estimate units.
5. Adjust for reality: Factor in price, lead time, channel, and region.
A simple version of step 4 looks like this:
# Rough demand estimate: baseline units scaled by an external trend index
baseline_units = 5000 # units sold for a comparable past style
trend_index = 1.30 # 30% rising signal from vision + search data
confidence = 0.75 # how much you trust the signal (0 to 1)
# Blend: don't bet the whole buy on the signal
forecast = baseline_units * (1 + (trend_index - 1) * confidence)
print(round(forecast)) # 5000 * (1 + 0.30*0.75) = 6125 unitsNotice the confidence factor. You never take an AI signal at face value. You weight it. A strong signal with low confidence should not blow up your buy.
AI forecasting is not equally reliable across the assortment.
Strong: Directional trends with clear visual attributes. Silhouettes (oversized, cropped), colors (a specific green becoming dominant), and recurring motifs. These show up in imagery and search with enough lead time to act.
Weak: Novel, never-seen-before items. Vision models learn from the past, so a genuinely new form has no training precedent. Also weak: micro-trends that spike and die in weeks, faster than your production lead time. If it takes you six months to make a garment and the trend lasts eight weeks, the signal is useless to you.
This is the core discipline: match signal speed to your supply chain speed. A fast-fashion player with a four-week lead time can chase micro-trends. A luxury house planning a year out needs slower, more durable signals.
Say you are a mid-market womenswear buyer planning next fall.
1. Pull the signal. Your vision tool shows oversized blazers in earth tones rising among your 25 to 40 audience. Google Trends confirms steady, non-seasonal search growth. Good, this looks durable, not a spike.
2. Find the analog. Last fall you sold a structured blazer. It moved 5,000 units at full price with light markdown. Use it as your baseline.
3. Apply the index with confidence weighting. Strong signal, decent confidence. You land around 6,000 to 6,500 units.
4. Split the buy. Do not commit everything upfront. Order a core quantity, hold open-to-buy budget (unspent buying budget kept in reserve) to chase if the trend accelerates.
5. Test and read early sales. Drop a small assortment early. Real sell-through in week one beats any forecast. Reorder the winners.
The AI does not replace the buyer. It gives the buyer a defensible starting number and an early warning.
🎬 [VIDEO: "How AI Predicts Fashion Trends" — youtube.com — a short explainer on how computer vision reads runway and social imagery to forecast demand]
Garbage in, garbage out. If your vision model is trained mostly on Western runways, it will miss signals in other markets. Know your training data's blind spots.
Correlation is not causation. A rising hashtag may reflect a viral moment, not durable demand. Cross-check across independent sources: vision, search, and your own early sales.
The self-fulfilling loop. If every brand buys the same AI-flagged trend, the market floods and everyone marks down together. An early signal is only an edge if you act before the crowd, or differentiate how you interpret it.
Overfitting to the tool. The forecast is an input, not an order. Keep human judgment in the loop, especially for brand identitybrand identityThe visual, verbal and cultural elements that define how your brand presents itself: logo, colours, tone of voice, and values.Voir la définition complète → decisions the data cannot see.
Vérification des acquis
1. Why does traditional forecasting (last year's sales plus merchant intuition) tend to fail specifically for trend-driven items?
2. What distinguishes a genuinely actionable AI trend-forecasting output from a vague observation like 'blazers are trending'?
3. In fashion, why is the timing of acting on a trend signal so consequential to the 'buy'?
4. Select ALL correct answers about how computer vision contributes to trend forecasting.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers describing the data layers AI trend forecasting combines.
Sélectionnez toutes les réponses correctes.
Most brands do not build vision models from scratch. They subscribe to platforms (Heuritech and similar providers) or blend a vendor signal with free tools like Google Trends and their own analytics team.
If you are non-technical, your job is not to code the model. It is to:
A forecast you cannot audit is a forecast you cannot trust. Every season, compare what the model predicted against what actually sold. That feedback loop is where the real value compounds.
AI will not make your forecasts perfect. Fashion is driven by culture, weather, celebrity, and chance, forces no model fully captures. What good AI forecasting does is shift the odds: earlier signals, fewer blind buys, tighter inventory, less markdown.
In a business where markdowns quietly destroy margin, moving the accuracy needle even modestly across a large assortment is worth real money.