Mr. Bets keeps an honest scoreboard.
Every pick logged before the event. Every result published after. No quiet edits.
On 1,024 held-out fights — the model has never seen these during training.
Picks at ≥65% model prob
Modern era, 2010–now
Per pick, automatic
Every one published
The board, model by model
68.8% on held-out · trained on 8,500+ fights
Weekly hit rate building below — no fake number
Pipeline in development, model launching this season
Atop the board after NBA ships
Mr. BetsMost pick services hide their hit rate. I publish mine. Both kinds — the wins and the ones I'd rather not talk about.
How this is computed
Every prediction we surface (dashboard recommendations, fight predictions, recommended bets) is logged at the moment it's shown. When the underlying event finishes we grade each one as a win, loss, or push and aggregate by sport over rolling time windows. No cherry-picking, no retroactive deletes — if a pick lost, it shows up here.
Beat the market — UFC
Cumulative units at a flat 1u stake per pick. Our model versus three baselines (always favorite, always underdog, coin flip). If the green line isn't climbing above the others, the model isn't earning its $9.99.
Calibration — UFC
When the model says 60%, does the pick actually win 60% of the time? Each dot is a 10%-wide probability bucket — its size reflects how many predictions fell in that bucket. The diagonal is perfect calibration.
UFC moneyline picks
Random Forest model, 22 features (Elo + Glicko-2 + strike/takedown accuracy + recent form). Predictions logged at the time they're shown to users; results graded after each event.
NFL picks (heuristic)
Heuristic — market-blended power rating, not a trained ML model. We publish its hit rate anyway so you can see what you're actually getting. A trained NFL pipeline is on the roadmap.
Convinced? $9.99.
The scoreboard keeps updating either way.
Subscribe and you get the dashboard, the picks, the tracker, and every future sport on the roadmap.