How the New Word Radar finds new words

Every 4 hours loud.fm scans nine radar lines — social, commerce, new domains and repos, registries, budgets, rules, replacements, cross-market and filings — to catch new words emerging in the US, Europe, Japan, Southeast Asia, the Middle East and Brazil, with sources, context and metrics.

Seven design principles

  1. Free signals find candidates; paid APIs only verify.
  2. Filters run in order of cost per elimination.
  3. Newness is per market: an old word in market A can be new in market B.
  4. Track parent words; expand child keywords from templates instead of paying to discover them.
  5. Fast markets are sensors for slow ones: meme tokens move in minutes, sites in days, inventory in weeks.
  6. Every result carries source links and context snippets.
  7. Every source can be switched on or off and budgeted on its own.

What a cycle does

The main cycle runs every 4 hours. Streams (newly certified domains, new on-chain tokens) flow into a buffer in real time; hot lists are pulled every 30–60 minutes; autocomplete, GitHub, Hugging Face, Product Hunt and Hacker News once per cycle; trademarks, SEC filings, the Federal Register, recall databases and concept boards daily. Each cycle freezes the buffer window, extracts and merges candidates, runs free filters, classifies and defines them with a model, verifies the top K within budget, scores them, updates their lifecycle and publishes a batch plus a change log.

Candidate extraction

jieba for Chinese, MeCab (fugashi) for Japanese, PyThaiNLP for Thai, wordninja for concatenated English domains, whitespace plus rules elsewhere. Latin-script languages use 1–3-grams with attention to casing and hyphenation; Chinese and Japanese use 2–6-character n-grams scored by cohesion (PMI) and branching entropy; structured sources such as trademarks, domains, token names and repo names contribute their names directly. Hashtags, quoted phrases, $tickers and the X in “what does X mean” are all recorded.

Nine sub-scores

novelty, velocity (Bayesian-smoothed growth), spread, consensus (new domains, tokens, trademarks), intent, demand (Trends vs anchor, volume, hot-list heat), competition (inverse SERP strength and supply), monetisation and lead — weighted per line into a 0–100 score.

Lifecycle

new → candidate (passed free filters) → watching (positive paid signal) → confirmed (multi-source, spreading) → cooling (growth turned negative) → archived (no signal for a long time; back to candidate if it returns); noise goes to rejected. Stages: origin (sentinel sources only), spreading (explanation demand or several platforms), mainstream (hot lists or steady search volume).

Compliance

Public data only; snippets of at most 200 characters with source links, never full text; terms of service and robots respected, unofficial endpoints kept low-frequency; only public account names, no profiling; the radar never trades.

FAQ

What is the New Word Radar?

loud.fm's New Word Radar is a data pipeline that runs every 4 hours. It reads public sources — social platforms, marketplace autocomplete, newly certified domains, GitHub and Hugging Face, trademark and SEC fund filings, procurement, regulations, recalls and shutdowns, filings and job posts — and surfaces words that are new and spreading in a given market, with sources, context, nine sub-scores and a lifecycle state.

How is “new” decided?

Newness is per market: an old word in market A can be new in market B. The baseline combines the wordfreq multilingual frequency lists, every word the radar has already seen, and brand/entity lists. Novelty, Bayesian-smoothed velocity, cross-source spread, consensus (new domains, tokens, trademarks), intent and more add up to a 0–100 score.

Where does the data come from?

Public data only: X, Telegram, Douyin, Xiaohongshu, Weibo, Bilibili, TikTok, Reddit, YouTube, Google and Amazon autocomplete, Certificate Transparency logs, GitHub, Hugging Face, Product Hunt, Hacker News, arXiv, SEC EDGAR, USPTO, FCC, Steam, Roblox, the Federal Register, the EU Official Journal, CPSC, openFDA, EU Safety Gate, GDELT, Chinese procurement portals, A-share concept boards, job boards and more. Every result links to its source; context is stored as snippets of at most 200 characters.

What can I see without an account?

The home page and public word pages show a few picks per cycle with a one-line definition. The full list, context, metrics, linked assets, verification results and history require sign-in; during the private beta, sign-up needs an invite code, and members can create API keys.

How do I use the API?

Create a key on your account page and send Authorization: Bearer <key>. Use /v1/changes to sync incrementally by seq, or /v1/words to filter by line, market, state and score. See the API page for details.

How often is it updated?

The main cycle runs every 4 hours by default. Streams (new domains, new tokens) enter the buffer in real time, hot lists refresh every 30–60 minutes, and registries and regulatory sources daily.