For twelve years, the panel on The Bitcoin Group has shaken a Magic 8-Ball, picked highs, called lows, and argued about where Bitcoin was going next. This is the running record — the ball, the panelists, the hits, the misses — pulled from 504 transcribed episodes against the daily close.
At the end of every show Thomas Hunt asks the ball: "will the price of Bitcoin be higher this time next week?" The verdict, scored against the seven-day BTC close on each episode, sits within a hair of even.
Each card links into the canonical TBG transcript mirror.
Every directional prediction we could attribute to a named TBG panelist, scored against the BTC close at the prediction's target window (next show, 30 days, 90 days, end-of-year, long term). Conditional ("if X then Y") calls are tagged separately and do not count toward hit rate.
| Panelist | Predictions | Hits | Misses | Inconclusive | Hit rate | Avg horizon | Last call |
|---|
The cumulative ball record, the by-year accuracy, the volume of panelist directional calls each year, the running hit rate of the top five panelists, and the BTC price line with markers for the highest-prediction-volume episodes.
Every TBG episode is auto-transcribed via Whisper and mirrored on this site at /tbg-mirrors/. There are no speaker labels in the raw transcripts — the model produces a single block of prose — so attribution to a panelist is done by proximity to a canonicalized name within the same sentence-window. If we cannot attribute a prediction to a confirmed-on-the-show panelist, it does not enter the leaderboard.
Names get garbled by Whisper. "Andreas Antonopoulos" appears as "Andreas Antonopolous", "Andreas A.", and other variants. "Jimmy Song" shows up as "Jimmie Song". We canonicalize each variant against a hand-curated roster using rapidfuzz token-set ratio plus Levenshtein, auto-merging at ≥ 75% confidence when first names match (≥ 85% otherwise) and flagging the rest for manual review at panelist_aliases_review.json.
Direction extraction uses a phrase library — "will go higher / lower", "I'm bullish / bearish", "this is a top", "target is $X", etc. Conditional sentences ("if X then Y", "depends on", "unless") are tagged and excluded from hit-rate math. Headline-read lines, questions, and "as X said" reportage are filtered out before extraction. Magic 8-Ball segments are routed to the dedicated tracker.
Hit / miss is judged against the BTC daily close at the target window. Default horizon is seven days unless the panelist explicitly named another window in the same sentence. If price data is missing for the target date, the prediction ships with a VERIFY badge instead of a hit/miss verdict. The 8-Ball record uses the same seven-day-close standard.
Every quote in this report is verified verbatim against the source transcript file before shipping — if the substring match fails, the prediction is dropped, not paraphrased. Source code lives in thunt-data-labs/collectors/tbg_predictions.py and tbg_magic8ball.py; intermediate data lives in verified-predictions.json and magic8ball.json.