Heuritech's algorithms are largely published open research and reproducible today. What is not reproducible — and where the real moat sits — is the proprietary 2,000-attribute labeled dataset, a decade of longitudinal trend history, curated social-media panels, and brand relationships. Replicating the tech is feasible; replicating the business is the hard part.
Founded 2013 by two ML PhDs (Tony Pinville, CEO) in France, now under Luxurynsight. Raised only ~$5.5M total, yet won the LVMH Innovation Award 2017 and counts Louis Vuitton, Dior, Moncler, and Havaianas as clients. That low-capex / high-signal ratio is the first clue: the moat is not compute.
The product: a SaaS platform (plus API) that quantifies and predicts what people wear by analyzing millions of social-media images daily, then forecasts trend trajectories up to 24 months out. The pipeline is four layers:
| Layer | What it does |
|---|---|
| 1. Data intake | ~3M social images/videos per day scraped from Instagram, TikTok, Weibo, Pinterest, YouTube. |
| 2. Panel construction | Segments accounts into Edgy (stylists, journalists, influencers — smallest, most predictive), Trendy (early adopters), and Mainstream (followers, lagging), plus geographic panels (Europe, US, China, India, Korea, Japan, Brazil, etc.). |
| 3. Computer vision | Proprietary detector for 2,000+ fashion attributes: colors, prints, fabrics, textures, shapes, silhouettes, necklines, sleeve types — even a brand's own product lines. |
| 4. Forecasting ensemble | 5 models blended by a master weighting (see below). |
The core insight: trend forecasting = watching what edgy accounts do before it hits the mainstream. Everything else is infrastructure around that signal.
The current ensemble is 5 specialized models, weighted dynamically per time series:
| Model | Type | Role |
|---|---|---|
| Next | HMM-RNN (in-house) | Detects disruptions via "early signals" from edgy influencers. The crown jewel. |
| PatchTST | Transformer | Best transformer method in literature (Mar 2023). |
| N-BEATS | RNN | Best RNN method in literature (Nov 2022). |
| ETS | Statistical | Classic exponential smoothing baseline. |
| SNAIVE | Statistical | Seasonal naive baseline. |
Key point: the R&D team published their own papers — Hermes (David et al. 2022, arXiv 2202.03224), the external-signals work (Springer 2023), and Next (HAL hal-04546265). PatchTST and N-BEATS are public literature. They claim ~90% accuracy and 12–24 month horizons, with a 5% MAPE reduction at 12 months from the new ensemble.
The forecasting algorithms are reproducible today. The data layer is not. Here's the split:
The tech is a thin wrapper around a data moat. If you try to beat Heuritech by building "the same thing, better algorithms," you're competing on the one axis that's already commoditized — and ignoring the axes where they have a 10-year head start.
Instagram deprecated public hashtag/tag scraping, TikTok and Pinterest aggressively rate-limit and block, and surviving APIs are gated or ToS-hostile. Scraping 3M images/day across 5+ platforms needs proxy infrastructure, anti-bot evasion, constant breakage patching, and carries real legal exposure (ToS violations, GDPR/CCPA on face data in images). Most clones die here, before writing a single model line.
"Capri" vs "cropped" vs "clam digger." "Chiffon" vs "organza." These are genuinely subtle visual distinctions. Public datasets (DeepFashion, iMaterialist) exist but are unbalanced, noisy, and don't cover Heuritech's taxonomy. You'd need a fashion-domain annotation pipeline — expensive, slow, and it's exactly the proprietary asset they spent a decade accumulating.
Forecasting needs years of past trend time-series to train and — critically — to validate. "Did our model correctly call the retro-football-sneaker surge in 2023?" requires that history. A new entrant starts at zero and can't shortcut 24-month forecasts without 5+ years of backfill.
Classifying "edgy" vs "trendy" requires fashion expertise and ongoing maintenance (accounts rise, fall, change). This is a standing operational cost, not a one-time model — and it's the piece most tied to their brand credibility.
Fashion trends are non-stationary and faddish; some live two weeks. "Early signal" extraction from sparse edgy-influencer activity is a hard cold-start problem. The HMM-RNN "Next" model is genuinely clever and non-trivial to get right even with the paper in hand.
LV and Dior buy from a company with a decade of track record and a known research team. A clone faces a brutal enterprise sales cycle where "we're 90% as good" isn't a sellable claim without a reference brand.
Skip trying to clone Heuritech head-on; you lose on the data moat. The winning play is asymmetry:
Heuritech's defensible edge is proprietary data, longitudinal history, curated panels, and brand trust — not model architecture. A challenger that acknowledges this and competes on a narrow, data-rich niche with foundation-model leverage has a real shot. A challenger that tries to out-algorithm them does not.