Hype, quantified
by AI.
See sneaker demand move before the market reacts. SoleSight fuses live search demand, real eBay resale asks, press coverage, social chatter and machine-learning 30-day forecasts into one explainable Hype Score per silhouette — rebuilt nightly by an autonomous AI pipeline, each signal labeled for exactly what it is.
AI consumer intelligence for sneaker brands, retailers and merchandising teams: spot rising silhouettes, time marketing, guide inventory bets with predictive analytics.
The Index
This week's movers
Launch radar
Demand events detected straight from the data — days where a model's search interest spiked to 3×+ its trailing 90-day baseline — plus where every model sits in its hype lifecycle right now.
Market intelligence
The analyst view — every brand and category in the index, rolled up: average hype, resale premium over retail, search momentum and share of the top 10.
Case study: catching a riser
Does it actually predict?
One case study is an anecdote. So the momentum engine — the Hype Score's heaviest input — is backtested across every model's full search history: thousands of historical checks asking one question, when a shoe was flagged rising, did that demand still hold weeks later?
Built for the decision-makers
The same index answers a different question for every desk that touches the sneaker market.
Brand merchandising
Which silhouettes have earned a retro or a colorway push — backed by demand data, not instinct.
Nike · Jordan · New BalanceRetail buying
Size buy quantities against live momentum and boutique sell-through instead of last season's gut feel.
Foot Locker · boutiquesMarketing teams
Time campaigns to demand spikes the Launch Radar catches as they form, not after they peak.
agencies · in-houseResale & marketplaces
Spot premium expansion early — acquisition and pricing signals straight from the ask side of the market.
StockX · GOAT · consignmentHow it's built
An end-to-end system: nine live data feeds, one autonomous nightly pipeline, zero servers — the platform publishes itself while its builder sleeps.
Collect
Google Trends, eBay (US · UK · DE), Bluesky, Mastodon, YouTube, Google News + sneaker press, Wikipedia and 15 boutique feeds — every adapter degrades gracefully if a source refuses.
Store
One SQLite database, committed back to the repo nightly — the commit history is a public, auditable heartbeat.
Understand
A RoBERTa transformer scores community sentiment; Prophet fits a 30-day demand forecast with confidence bands per model.
Detect
Launch Radar flags demand spikes ≥3× each model's trailing baseline and labels every lifecycle stage.
Score & explain
Five weighted signals blend into a 0–100 Hype Score, plus a plain-English insight per model.
Publish
Everything lands in one static data.json; GitHub
Actions redeploys this site — no human in the loop.
Tool choices, briefly: Prophet because sneaker demand is seasonal and per-model history is short · RoBERTa because community chatter is noisy short-form text · SQLite + one static JSON because zero infrastructure is a feature, not a compromise.
The boutique network Kith · Undefeated · A Ma Manière · Union LA · Extra Butter · Nice Kicks · Packer · Notre · Xhibition · Bodega · Concepts · Feature · Lapstone & Hammer · Social Status · Sneaker Politics — shelf availability polled nightly from each store's public product feed.
Signal provenance — every number, traced
The Hype Score is five weighted signals. Here's exactly where each one comes from and how live it is — nothing black-box.
resale ← resale.py
live
trends ← google_trends.py
live
trends
live
social ← bluesky.py, social.py
partial
reddit_posts ← sentiment.py
live
Context signals shown across the site — Wikipedia attention, press coverage & momentum, boutique sell-through — stay out of the score until the backtest earns them a weight.
How the score works
Each signal is labeled for what it is today — live real data refreshed nightly, api-ready real adapters awaiting free API keys (demo data until then), or partial some platforms real, the rest modeled, or modeled synthesized from real search momentum pending a viable API.
Resale premium
Median deadstock ask (new, unworn) vs. retail from live eBay listings, nightly — the market's dollar vote.
Search momentum
Google Trends interest, last 14 days vs. the prior 14 — movement within each model's own history.
Social buzz
Bluesky and YouTube live; Instagram/TikTok modeled (no viable APIs).
Search intensity
Current interest relative to that model's own historical peak — never compared across models.
Community mood
Transformer sentiment over live Bluesky, Mastodon, Reddit & YouTube-comment chatter — four keyless community sources.
Technical methodology — formula, normalization & a worked example
The formula
score = Σ(wᵢ · vᵢ) / Σ(wᵢ) over available signals
Every component maps onto 0–100 before weighting:
- Search intensity — Google's 0–100 index, 14-day mean. Relative to each model's own peak, so it measures how hot the model is vs. itself, not raw volume across models.
- Search momentum — recent 14d vs. prior 14d percent change, mapped as
50 + Δ% × 0.8, clamped to [0, 100]. Saturates at roughly ±62%. - Resale premium — median price ÷ retail MSRP, mapped linearly from 0.8× (→ 0) to 2.6× (→ 100). The clamp caps extreme listings; prices are decile-trimmed before the median to drop fakes and typo listings. A model with no known retail price contributes nothing (see missing signals).
- Social buzz — daily cross-platform engagement, normalized 0–100 to the model's own peak (same convention as Trends).
- Community mood — per-post transformer sentiment
P(pos) − P(neg)∈ [−1, 1], averaged over all scored posts, mapped to 0–100.
Missing signals: weights renormalize over whatever is present — a model missing resale data is scored fairly on the rest, never implicitly zeroed. Update cadence: search & forecasts nightly (stalest-first); other signals refresh whenever their source runs. Baseline window: each model carries ~269 days of daily history.
Worked example — today's #1
Known limitations — stated, not hidden
- Resale reads asking prices, not completed sales — sold-price feeds are partner-gated. Asks run slightly above realized sales, and the decile trim only softens that.
- Google Trends is relative interest within each model's own history, never absolute volume — no score ever compares raw search volume across models.
- Community mood reads Bluesky, a smaller pond than Reddit or X; Reddit joins with a free key, X's API now bills per read.
- Instagram/TikTok buzz is modeled from real search momentum — neither platform offers a viable free API, and the card says so.
- Wikipedia attention, press coverage and boutique sell-through stay outside the Hype Score until a backtest earns them weights.