TG GiftsPublic changelog — method, data, dataset
Last change:
- — Dataset: version 2026.11.03 is published on Zenodo with a DOI, 10.5281/zenodo.23897620, licence CC BY 4.0, six files. floors-daily.csv holds 1,613 rows across 102 collections and 16 days; hunts.csv holds 35 hunts. Same numbers, now with a permanent identifier.
- — Data and page: the cross-market floor comparison goes live. Each collection page and /floors show the cheapest floor another market lists for the same collection, with the market name, the time it was read and a link to the listing, sorted by the size of the gap. It adds a comparison and changes no floor, weight or index value; the index stays Fragment-only.
- — Correction: the public changelog said a 5.2% median error for usernames. That figure was in-sample, measured on the same 90 days the v123 parameters were tuned on. It was replaced by the out-of-sample figure: 28.2% median error (95% interval 21.8–42.8%), n=170, preliminary, with 37.1% of estimates within ±20%. Old value 5.2% (in-sample, n=2,140); new value 28.2% (holdout, n=170).
- — Dataset: release v2026.11.01, a tagged snapshot of the open data. Affects no number.
- — Dataset: GitHub release v1.0.0, a tagged snapshot of daily gift floors plus the TG Gifts Index. Affects no number.
- — Dataset: release v2026.10.29, the first tagged release of the open data, CC BY 4.0, with CITATION.cff, datapackage.json (Frictionless), floors-daily.csv and weekly files. Same numbers, now citable.
- — Model v123, evidence-based class values: the username class value becomes a recency-weighted median (half-life 60 days) of cleaned sales instead of a flat 12-month median; patterns are compared only with their own shape; round trips, one seller's identical-price batches, far-from-market prices and on-chain-contradicted sales are excluded with a counter; the range widens to the 5th–95th percentile. Affects every username estimate; walk-forward backtest published on 2,140 real Fragment sales from the last 90 days.
- — Model v136: every floor, below-floor and value label states that it is a Fragment floor and when it was checked (site in 8 languages, bot in EN/RU/FA, API field floor_source); an estimate equal to the Fragment model floor is capped at medium confidence. Labels and confidence only; no estimate changed.
- — Model v118, valuation honesty: gift range always ordered, confidence label added, no verdict on thin markets, recent sales read newest first, a relisted handle keeps its own last sale. Ranges and confidence labels change; no numeric benchmark, because these are rule changes.
- — Claim: all pages, datasets and posts are authored as TG Gifts Team, with no personal names published. Affects no number.
- — Data: first weekly market report (2026-W43), covering 2026-10-19 to 2026-10-25 and 101 collections, frozen after publication and changeable only through a visible corrections section. New product, same source.
- — Dataset: first per-model daily file, models/2026-10-19.csv, one row per model with rarity_permille and buy_now_listings; the GitHub repository w3radar/telegram-gift-floors is created. Adds a second level of granularity; collection floors unchanged.
- — Method: index v1.0 and base 100. Universe, daily snapshot, eligibility (a floor on both days and at least 15 recent sales), weights proportional to recent sales capped at 10% per collection, chain-linked formula, freezing rule. This is the origin of the series.
- — Data: first daily floor point, one per collection per UTC day, and the first public hunt, logged at 23:30 UTC. This starts floors-daily and the track record.
- — Data: the first real-data scanner reading public Fragment pages goes live. It creates the numbers; there is no earlier history to recover.