{"source":"https://tggifts.com/how-we-value","method":"Every day a stratified sample of our own estimates is frozen, then compared with the real sale price when the item sells. Error = |estimate - sale| / sale.","public":false,"settled":9,"open":141,"window_days":90,"updated":null,"note":"Collecting: numbers are published once 30 estimates have settled against real sales.","model_changelog":[{"v":"v123","date":"2026-10-28","title":"Evidence-based class values, suspicious-sale filter, on-chain check","changes":["Username class value is a recency-weighted median (half-life 60 days) of cleaned sales instead of a flat 12-month median","Pattern handles are compared with their own shape (blocks like aaabbb vs interleaved like ababab)","Round trips, one seller's identical-price batches, far-from-market prices and sales the TON blockchain contradicts are excluded, with a counter; wallet-link signals are only counted","Range is the 5th to 95th percentile (it was the 25th to 75th), widened by 1 GRAM granularity under 20 GRAM","Sales used as evidence are listed with date and a Fragment link; on-chain verification ticks where checked","Daily accuracy monitor: every estimate is frozen, then compared with the real sale price"],"before":null,"after":null,"note":"walk-forward on 2,140 real Fragment sales, last 90 days; ranges got wider (median high/low 5.0x to 7.8x), which is the price of honest coverage","in_sample":{"n":2140,"medianError":0.052,"coverage":0.805,"window":"days 0-90 before the v123 change (the parameters were tuned on this window)","note":"in-sample: optimistic, not a forecast; kept only for transparency","before_medianError":0.15,"width":7.8},"backtest_out_of_sample":{"label":"BACKTEST (historical sales), not the live track record","segment":"usernames","n":170,"medianError":0.282,"ci95":[0.218,0.428],"coverage":0.759,"preliminary":true,"window":"days 90-180 before the v123 change (holdout: never used for tuning)","method":"rolling-origin walk-forward on real Fragment sales; bootstrap 95% interval (400 resamples)"},"live_track_record":"fields settled / overall / segments of this file (published after 30 settled estimates)"},{"v":"v118","date":"2026-10-27","title":"Valuation honesty","changes":["Gift range is always ordered; confidence label; no verdict on thin markets; recent sales read newest first; a relisted handle keeps its own last sale"],"before":null,"after":null,"note":"no numeric benchmark (rule changes)"},{"v":"v136","date":"2026-10-28","title":"Floor source and time on every floor label","changes":["Every floor, “below floor” and value label now says it is a Fragment floor and when it was checked (site in 8 languages, bot in EN/RU/FA, API field floor_source).","A live check found other gift markets list the same collections a median 13% lower than Fragment (37 of 101 collections more than 15% lower). Floors stay Fragment-only for now, with a visible note that other markets can list lower.","An estimate that equals the Fragment model floor (no sales cap) is shown with at most medium confidence."],"note":"Labelling and confidence change only: no estimate number changed."}]}