Trend and demand sensing from search, social, and early POS signals

# Trend and demand sensing from search, social, and early POS signals

A buyer at a fast-fashion label commits to 40,000 units of a ribbed knit cardigan six weeks before it hits stores. By the time she sees the first sales report, the factory slot is gone and the color assortment is locked. If the buy is wrong, the company eats markdowns. If it is right by luck, they sell out and leave money on the table because they cannot reorder in time.

Demand sensing is the discipline of shrinking that guessing window. Instead of relying on last season's numbers and gut feel, you blend live signals (what people search, what they share, what actually rings up in the first two weeks) to correct the forecast before it is too late.

This lesson shows you how to build that blended view.

What "demand sensing" actually means

Demand forecasting predicts future sales weeks or months out, usually from historical patterns. Demand sensing is shorter-horizon: it uses fresh, high-frequency signals to adjust that forecast in near real time.

In apparel the payoff is huge because product lifecycles are short and mistakes are expensive. A missed trend means dead stock and markdowns (selling below full price to clear inventory). A missed winner means lost sales you cannot recapture.

Three signal families do most of the work:

• Search intent: what people type into Google before they buy.
• Social velocity: how fast a style, color, or silhouette is spreading.
• Early point-of-sale (POS): what actually sells in the first days of a drop.

Each one is noisy alone. Blended, they are far more reliable.

Signal 1: Search intent from Google Trends

Google Trends is free and shows relative search interest over time and by region. It will not give you absolute volume, only an index from 0 to 100, but the shape of the curve is what matters.

Practical uses:

• Rising queries: compare "quiet luxury" versus "coquette" versus "barrel jeans" to see which silhouette is gaining.
• Seasonality: "linen dress" spikes every spring. Knowing the exact ramp week helps you time the drop.
• Geography: a term heating up in Seoul or London often precedes a US wave.

The key move is reading the slope, not the level. A style at index 30 and climbing 15 points a week is more interesting than one sitting flat at 80.

A caution: search interest is not demand. People search "Y2K fashion" for many reasons besides buying. Treat it as a leading indicator, not a sales number.

Signal 2: Social velocity

Social platforms show what is spreading and how fast. The useful metric is velocity: the rate of change in mentions, saves, or views, not the raw total.

What to track:

• Hashtag and sound growth on short-video platforms, where a trending audio can carry a garment style overnight.
• Save and share rates on visual platforms, which signal purchase intent better than likes.
• Creator adoption: when mid-tier creators (not just mega-influencers) start wearing a silhouette, mainstream demand often follows.

Be disciplined about what you collect. Scraping personal data or violating platform terms creates legal and reputational risk. Use official APIs, licensed social-listening tools, or publicly reported trend data, and check platform terms before pulling anything.

🎬 [VIDEO: "How Zara Uses Data To Dominate Fashion" — youtube.com — a clear walkthrough of fast-fashion's data-drivendata-drivenAn approach where decisions are systematically informed by data analysis rather than intuition alone.Voir la définition complète →, short-cycle model]

Signal 3: Early POS (the first two weeks)

Point-of-sale data is the ground truth: real money, real units. The problem is timing. By the time full sales data arrives, the reorder window is closed.

Demand sensing solves this with early read: use the first days of sell-through to project the full run.

Sell-through rate is the share of received inventory sold in a period. If you shipped 5,000 units to stores and sold 1,200 in week one, that is 24 percent sell-through. Early sell-through is a strong predictor of full-lifecycle performance, especially when you compare it to a benchmark curve from similar past items.

Two moves make early POS reliable:

1. Normalize by exposure. A style in 200 stores will show more units than one in 40, even if the second is hotter per store. Always look at units per store per day, or sell-through percent, not raw units.

2. Compare to a reference curve. Ask: is this item ahead of or behind the typical week-one pace for its category? Being 30 percent ahead of curve is your signal to chase (reorder more).

Blending the three signals

No single signal is trustworthy. The art is combining them so they check each other.

A simple, transparent approach: convert each signal to a normalized score, then weight them by how much lead time and reliability they offer.

• Search: earliest signal, moderate reliability.
• Social: early, high noise, strong on emerging aesthetics.
• Early POS: latest but most reliable, because it is real sales.

Here is a minimal illustration of the logic (not a production model):

# Normalize each signal to a 0-1 score, then blend.
# Weights reflect reliability + lead time, and must be tuned to your data.

def demand_score(search_idx, social_velocity, sellthrough_vs_benchmark):
    # search_idx: Google Trends slope, scaled 0-1
    # social_velocity: week-over-week mention growth, scaled 0-1
    # sellthrough_vs_benchmark: actual / expected week-1 sell-through, capped
    weights = {"search": 0.25, "social": 0.30, "pos": 0.45}
    score = (
        weights["search"] * search_idx
        + weights["social"] * social_velocity
        + weights["pos"] * min(sellthrough_vs_benchmark, 1.5) / 1.5
    )
    return round(score, 3)

# Example: strong social, POS running ahead of plan
print(demand_score(0.4, 0.8, 1.3))  # -> a "chase" candidate

Before any POS exists, the model leans on search and social. Once week-one sales land, POS dominates because it is real. This is the whole point: the blend shifts weight toward harder evidence as it arrives.

Two guardrails:

• Backtest before you trust it. Run the model against past drops and check whether high scores actually predicted winners. If they did not, the weights are wrong.
• Keep a human in the loop. A buyer's category knowledge catches things data misses, like a garment that photographs well but fits poorly.

Turning the score into a decision

A score is useless unless it drives an action. MapMapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.Voir la définition complète → scores to concrete moves within the constraints of your supply chain:

• High score, ahead of curve: chase. Place a reorder if lead time allows, or shift open-to-buy budget toward this style.
• Medium score: hold and monitor. Wait for more POS days before committing.
• Low score, behind curve: mark down early or reallocate stock to stores where it sells better.

Speed matters more than precision here. A rough signal acted on in week one beats a perfect signal in week five, because the factory slot and the selling season wait for no one.

Open-to-buy is the budget a buyer has left to spend on new inventory in a period. Demand sensing helps you spend it on proven early winners rather than locking it all up before any signal exists. Many fast-fashion operators deliberately hold back a portion of the buy to chase in-season, and reliable early reads are what make that holdback profitable.

Common traps