How we predict — and how accurate we've been
59.0%
Model accuracy
on 156 historical picks
+9pp
Edge over 50/50
vs coin-flip baseline
424
Bouts backtested
point-in-time, no leakage
The PropsBot model is a transparent factor-based scoring system — not a black box. For every bout we score the two fighters on ten+ dimensions, combine via weighted sum, and convert to a probability via a logistic function. That model probability is then compared to the de-vigged moneyline from the market to flag potential value.
BKFC win rate (Bayesian-smoothed)
18%
Recent form (last 5, decay 0.85ⁱ)
14%
BKFC KO rate / finishing ability
13%
Age curve (steep decline past 35)
12%
Reach advantage
10%
Style matchup (pressure/counter/etc)
8%
MMA-background depth
8%
Activity / ring rust
7%
Cut susceptibility (past DS losses)
6%
Strength of schedule
5%
BKFC experience (fight count)
5%
Height advantage
4%
Per-factor accuracy from our point-in-time backtest (skipped where data is thin):
Recent Form (n=122)
58.2%
BKFC Record (n=132)
56.1%
Height (n=125)
52.0%
Reach (n=152)
52.0%
Age (n=175)
50.9%
Each factor's edge strength is its own sub-score — a 1" reach gap barely moves the needle; a 6" gap is meaningful. Missing data drops a factor's weight to zero rather than guessing.
Be honest with yourself: 59.0% accuracy on toss-up-filtered picks means we're wrong roughly 4 out of every 10 bets we recommend. The model is a decision aid, not a crystal ball. BKFC intangibles (cut depth, weight cut, camp news, injuries) aren't fully captured. Never bet more than you can lose on a pick without confirming with additional research.
