Methodology
Last reviewed · Editor Wan Meng (/en/about)
Verdicts, resale scores, and future-value forecasts are produced from structured observation plus statistical models — not LLM inference. This page documents the underlying math, data sources, and error bands. The full API reference is public at /en/docs.
How is the verdict computed?
Asking price is compared against the same-SKU P10/P50/P90 distribution. ≤ P30 is deal, P30–P70 is fair, ≥ P80 is overpriced. For handbags we anchor on resale median rather than retail (because Hermès quota and gray-market pricing make retail noisy). Endpoints are public at /docs.
Where does price data come from?
New retail: brand boutiques and 6 affiliate retailers (Farfetch, Mytheresa, NAP, Saks, Bergdorf, Selfridges) via authorized feeds. Resale: 7 marketplaces (Vestiaire, TRR, Fashionphile, Secoo, Plum, PangHu, ZhiEr) via affiliate plus public-page observation. We do not scrape protected content.
What is a field-level provenance badge?
Every numeric value carries a colored dot: green = manual (human verified), blue = real sample, yellow = llm_synth (AI-synthesized, dated), gray = null (no observation). This is the gap Phia left after Fortune flagged data accuracy issues — we annotate provenance per field instead of per page.
What error bands apply to resale median and 3-year forecast?
Resale median: ±10% confidence interval (sample n ≥ 30). 3-year forecast: ±15% based on a brand × style × liquidity 3-factor model. Verdicts include sample size n, and any output with n < 30 carries a low_confidence badge.
Why don't you let an LLM produce the price directly?
An LLM cannot know what a Birkin 25 traded for on Vestiaire yesterday — it can only statistical-extrapolate from training data, which freezes stale information into outputs. Our boundary: LLMs only do SKU normalization (multilingual embeddings to align 'Black Togo Birkin' with 'Hermès BK25 Black Togo'); all price data flows from structured observation.
How do you handle FX?
Monthly mean from ECB and IRS rates. Cross-currency comparisons display original currency plus USD conversion. Live FX is intentionally excluded — minute-by-minute fluctuation makes verdicts flicker without giving the user useful signal.
Do you authenticate? How are counterfeit risks flagged?
We do not perform final authentication. We flag risk and refer out: when a listing's price, seller, or description deviates statistically from same-SKU history, we attach a risk badge. Authentication is referred to Entrupy, CSA, Real Authentication, or the brand's own service.
Where are the API docs?
Public OpenAPI: https://luxecopilot.com/en/docs (static summary + curl examples + endpoint index). All endpoints (verdict, fair-price, resale-score, future-value, wardrobe-roi, lens) expose schema and examples. Live OpenAPI under migration; /en/docs is the recommended entry today.
Spotted an error or have data to add: hello@luxecopilot.com. Each Journal article repeats observation window and sample size inline (/en/journal).