Upcoming fights
Who we think wins each fight, and the few fights where the odds are good enough to bet on.
About these numbers
How did we do?
Card by card
Every card
Return is how much each card moved the bankroll under the chosen system. ROI is profit divided by the amount staked on that card. Draws and no-contests are left out.
Choose your betting system
Pick how much risk you want to carry. The stats, this week's bets, the track record and the rules below all follow your choice. Results are from 2021 onward. Each system's settings were chosen using 2012–2020 only, but the 2021+ results were visible to us while the system families were designed, so treat them as somewhat optimistic.
This week's bets
Your bankroll figures stay in this browser. Stakes are what the chosen system would place at the best listed price; take the bet only at that price or better.
Track record
How the system bets
What to expect
Advanced analysisCompare every system · 1,512-variant explorer · edge sweeps · bet ledger
Systems compared
Holdout 2021 → today
Calmar = CAGR ÷ max drawdown. Ulcer index = root-mean-square drawdown over time, a measure of how deep and how long you sit below a previous peak. Lower is easier to live with.
Selected strategy
Bankroll · 100u compounding
dev period → untouched holdout
Season by season
Edge threshold sweep
flat-stake ROI, other settings fixed
When to bet
same rules, different price source · holdout
CLV = ln(price taken ÷ fair closing price). Positive mean CLV over hundreds of bets is the fastest reliable sign of real edge.
Strategy explorer
Ranked on the development period only. Holdout columns show what happened next, and that is the number to trust. Click or press Enter on a row to load it.
Latest holdout bets
last 60 bets by the selected strategy
How good are the predictions?
Every number here comes from fights the model hadn't seen: each season was predicted using only earlier fights. The yardstick is the bookmakers' closing odds, the sharpest public forecast available. Figures cover 2021 to today.
How often the pick wins
Only pick fights where the favourite's chance is at least the value on the axis
Confidence tiers
Model + market, 2021 → today
Are the percentages honest?
What the model looks at
Share of the model's decisions driven by each input
Accuracy season by season
Share of fights where the pick won
Technical evaluationLog loss ladder · calibration statistics · blend weights · method-of-victory model · subgroups
Walk-forward out-of-sample metrics. Development = 2012–2020 (used for choices); holdout = 2021 onward.
Baseline ladder
bouts with a closing line · lower log loss is better · skill > 0 beats the market
Confident-pick tiers by source
A high win rate comes mostly from picking favourites, so the ROI column shows what backing every pick at the best available price returned.
Log loss by year
lower is better
How much the blend trusts the model
logit(p) = βm·model + βk·market
βm above zero means the model carries information the closing line does not.
Method of victory
KO/TKO · SUB · DEC given the winner · walk-forward vs division base rates
Stability by subgroup
log loss on bouts with a closing line, all out-of-sample years
Does the referee change the result?
Every UFC referee since 1997, checked against every bout we hold. A referee can't score a fight, but decides when it ends and when grappling gets stood up. Each referee is compared with what the two fighters' records predicted, and with the other referees who worked the same card, so a referee who simply gets the heavy hitters isn't mistaken for a quick stopper.
What we tested
Referees compared only with others working the same card · 2,000 reshuffles each
Referee by referee
Profiles
Points above (+) or below (−) what the fighters predicted · era-adjusted · referees with 50+ bouts
Grey = well within chance (|z| < 2). A dot marks a gap that stays significant after correcting for testing many referees at once. Ground control is the share of fight time spent in a controlling position; upsets are underdog wins against what the closing odds implied.
Does it improve the predictions?
Change in log loss on 2012–2020 fights · negative = better · 95% range from reshuffling whole events
Who referees where
Referee assignments are only confirmed on fight night. What is known when a fight is booked is the commission, and each commission uses a small pool. Last three years.
What could a season look like?
10,000 possible seasons
Odds of each outcome
Over one season, starting from 100%
How seasons end
If the edge is smaller than we think
Past results usually overstate the future. Here is the same system under weaker assumptions.
Your safety rules
Decide these before you start, and don't change them after a bad run
How the simulation worksMethod, assumptions and exposure limits
Method
Exposure limits
How AlgoMMA makes its picks
A short guide to what the system does, how we make sure its results are real, and what it can't do.
How we keep it honest
Where the data comes from
Limits you should know
Bet responsibly
Technical detailsPipeline · automated integrity tests · odds coverage by year
Pipeline
ufcp all · build → train → test → predict → report
Automated integrity tests
Odds coverage
Keep AlgoMMA free. No paywall, no ads. If it helps your fight nights, chip in to keep the models training and the odds fresh.
Support on Ko-fi