Model Lab.
Explore a matchup. Adjust the influence. Compare the evidence.
Working research goals, not established performance benchmarks. Compare accuracy and probability scores on the same fights; confirm improvements with future pre-event predictions.
Can a fighter overcome their reach disadvantage? (Archived v4 study)
This test uses earlier fights to measure performance above expectations against opponents with a similar reach gap. A general-form control checks whether it is finding reach-specific ability or simply recent overperformance.
| Research configuration | Winner accuracy | Correct picks | Log loss ↓ | Brier ↓ |
|---|
How the reach-history test works
Expected results come from models trained before each fight year, with a pre-bout Elo warmup for the earliest archive. Histories update after the entire fight day. No future outcomes are used for earlier profiles. These are reused historical comparisons, not fresh confirmation. This research panel does not change the active model or its controls.
Current inputs. Same fitted model.
Loading input refresh status…
Completed main-UFC events are replayed automatically from saved validated results, totals and rounds. Same-day results are held until the UTC day closes. No coefficient retraining occurs. Missing or invalid stats add no evidence; coverage is shown here. Historical comparison predictions remain frozen. Corrections before the seed cutoff require a new audited seed; missing target/position splits await valid source data.
See the weights. Test the change.
Adjust a group or an individual input. 0× removes its contribution; 1× uses the original fitted influence; 2× doubles it. These are coefficient multipliers, not percentages of total model importance. A negative fitted coefficient can reflect overlap with other inputs; its sign or size alone does not measure fighter skill.
The comparison below isolates weight edits. Hometown is measured separately in its own panel. Reused historical results can overfit; “Improved” describes this archive, not proven future performance.
Flexible or conservative?
Regularization pulls learned coefficients toward zero during training. Smaller C means stronger shrinkage. It refits all inputs together; it is different from scaling a group after training.
Loading completed, real offline refits after unlock. This button compares saved trained models and saves the result; it does not launch fresh training in your browser. Weight-slider previews are independent of these fits.
How individual coefficients changed
Coefficients use different input units. Their raw sizes are not a ranking of importance. Both the ordinary and unknown-data coefficient are shown.
| Input | Current coefficient | Refitted coefficient | Current missingness | Refitted missingness |
|---|
Saved training comparisons
Every input. Every rule.
Current fitted settings, data rules and experiment boundaries in one place. Labels show what you can change instantly, what has a saved policy, and what needs training or a historical rebuild.
Full fitted specification & style transform
Source-backed settings and exact transform parameters. Per-input coefficients, training medians, scales and unknown-data terms are available in the input editor below. Rebuild-only rules are read only here; changing their text would not change the model.
Test the home advantage.
A separate, learned adjustment to the base prediction. Compare both versions on the same fights. Birthplace is a proxy for hometown; it is not residence or training location.
Learned from the difference between earlier outcomes and the base model’s expectations, using a three-year window and strong shrinkage toward zero. The fitted effect can be negative. Half applies half the learned log-odds effect; it is not a fixed percentage-point bonus. Off returns the base probability exactly.
Location coverage and historical research
The verified archive has 60 eligible fights within 25 km, with 37 hometown wins (61.7%). The earlier reported 74-fight sample has no recoverable matching table and is superseded for implementation. GeoNames cities15000 city centers exclude some small towns. Unresolved places stay unknown. This archive has been examined repeatedly and is not fresh validation.
Location attribution: GeoNames · CC BY 4.0. Source geography refreshes from saved fighter profiles and event venues. New places outside the verified registry need review before they can qualify.
Has it helped on new fights?
Both predictions are frozen once, during the 48 hours before the published event start, when a saved card was checked within 24 hours. Forecasts use automatically rebuilt performance inputs and current Elo with the same approved fitted coefficients. Their input generation and history cutoff are retained. Logging pauses while required profile updates are being prepared. The hourly refresh logs forecasts and checks results automatically.
| Frozen matchup | Base A | Adjusted A | Policy | Snapshot dates | Status |
|---|
Switching Off affects new forecasts only. It never deletes or rewrites frozen forecasts. Removed/replaced bouts, draws, no contests and unverified results are excluded from scoring. Historical previews and official saved picks/odds remain separate.
What drives the pick?
Contribution bars show base-model log odds. The separate hometown adjustment is shown above.
Contribution to this prediction
Left / teal favours B. Right / pink favours A. Values add up in log odds; they are not percentage-point bonuses or causes.
Inspect the strongest individual inputs
Every input. Every weight.
114 prepared inputs, each with its own fitted coefficient and a separate missing-data coefficient. Nothing is hidden in a top-ten list. Use 0× to remove a term, 1× for the original fitted value, or up to 2× to increase its influence. Category multipliers also apply.
| Model input / explanation | Fitted coefficient | Input multiplier | Unknown-data multiplier | Main contribution | Experiment contribution |
|---|
Prediction = sigmoid(sum of contributions). There is no fitted intercept; equal profiles start at 50%. Coefficients are per stored input unit, so their sizes are not comparable importance scores. Contributions include the selected matchup’s actual difference. The unknown-data multiplier changes only the learned missingness signal; median imputation remains. A zero fitted coefficient stays zero when multiplied.
How these inputs are built
These upstream settings are shown for completeness. Changing them needs a historical rebuild and retraining; the live controls change existing contributions only.
Does the change help?
Evaluate the same saved comparison fights with the proposed contributions. This is a sensitivity test, not retraining or fresh validation.
Main settings loaded. Run a comparison to inspect the breakdown.
| Configuration | Winner accuracy | Correct | Log loss ↓ | Brier ↓ | Changed picks |
|---|
These historical outcomes have been examined repeatedly. A better score here can overfit; confirm promising settings on later, unseen fights before promotion.
Which changed picks helped or hurt?
| Date / event | Matchup | Base A | Experiment A | Official winner | Effect |
|---|
Results by year
Results by division
Results by prior UFC experience
Keep the useful ideas.
Saved settings and results are private and versioned. Load restores a configuration; it does not promote it.
Elo: competition & recent form
Compare four completed experiments. Stronger competition sensitivity uses a 300-point expectation scale instead of 400. Form adds up to 80 × an opponent-adjusted recent-results score, with 50% weight on the latest result and a one-year decay half-life.
These are saved historical experiments, not live settings. No candidate is promoted. The 2022–2026 archive has been examined repeatedly; small gains are exploratory. Offline title coverage matches the approved historical model and is less complete than live Elo.
Khabib: points earned per fight
Same offline data and title coverage for every version. Ratings shown here are research snapshots, not the live title-enriched leaderboard.
Changed picks: helped / hurt
Full-model refits on the 2,484-fight comparison archive.
Exact experiment rules and limitations
Stronger form & losses to elite opponents
Sixteen saved experiments compare form strength, opponent sensitivity, title bonuses and finish bonuses. Saved title metadata covers all 8,591 historical UFC inputs, with 439 eligible title bouts. The approved archive remains unchanged.
The title-enriched research reference includes saved title flags absent from the approved historical inputs. Compare rule changes against that reference as well as the approved archive. Outside entry calibration remains fixed, so the research reference is not an exact reconstruction of every live rating.
All results below use the same 2,484 repeatedly inspected 2022–2026 fights. C=.01 stays fixed; other prepared inputs stay unchanged. These are exploratory comparisons.
Khabib: Horcher and Poirier gains
Volkanovski: fight-by-fight losses
Full replays include earlier effects of each rule. “Expected” is Elo’s own pre-fight expectation, not the full prediction model.
Remove bonuses while holding pre-fight ratings fixed
Rating inflation from loss-only discounts
Reducing only the loser’s deduction creates points. Symmetric bonus removal preserves equal transfers. No recentering is applied in these experiments.
Changed picks versus title-enriched reference
Earlier diagnostic scores and exact rules
Earlier annual logistic fits retain approved prepared transforms, including division styles fitted through 2021. They are diagnostic checks, not fully rebuilt independent validation.
Odds for newcomers or extreme favourites
Compare sparse-history Elo updates, provisional newcomer ratings, and updates triggered only by a large market favourite. Extreme-only rules ignore close-market prices for that bout’s update; earlier market influence can persist in later ratings.
Research only. Strict pre-event prices are the primary evidence; event-start snapshots are a separate sensitivity check. Provider timestamps are not independently verified. There are zero eligible 24-hour-before-event prices in the comparison period, so these results cannot establish performance at our earlier pick deadline or a betting edge.
Compare with the market on identical fights
Only fights with eligible current-bout prices. The market benchmark uses later information than an earlier published pick may have had. No profit or ROI is inferred.
Khabib’s Horcher win
The saved Horcher price is stamped at event start. It is excluded from the strict primary policy. Inclusive results below are illustrative sensitivity results, not independently verified pre-event evidence.
Provisional entry adjustments
Entry priors affect starting ratings and may affect that bout’s prediction. Update-only rules first act after the result. Entry shifts are separately measured; all result transfers remain equal and opposite.
Changed picks: refits versus independent baseline
Exact rules and coverage audit
Kalshi: prices saved before the deadline
The existing hourly refresh collects locally timestamped fight prices. At 24 hours before the advertised event start, the latest pre-deadline observation is frozen. Compare the independent core, a fixed-weight UFC debut-entry preview, and Kalshi on the same future fights.
This is a new prospective cohort, separate from historical accuracy. The preview uses the approved coefficients with a 50% provisional Elo correction capped at ±200 points for debutants. It does not change stored live Elo, carry market adjustments into future careers, apply the hometown overlay, or publish picks.
Latest observations
All frozen, scored model forecasts
Identical fights with eligible Kalshi prices
Matched UFC debutant fights
Frozen forecasts and exclusions
Fixed quality rules, policy and limitations
Young-favourite confidence
Under25 on event day · exactly1 or2 earlier UFC appearances · already favoured by the independent model. Tested coefficient0.2210917212; no odds, Elo changes or automatic retuning. Applied after hometown; skipped if hometown changed the favourite. Picks stay identical.
Policy changes affect new predictions. Frozen forecasts retain their original policy. Compare eligible fights using log loss and Brier (lower is better); accuracy stays identical. Small samples and repeated checks do not establish a benefit. No historical scores are mixed into prospective results.
Frozen official forecasts · original and adjusted
Policy, calibration bins and reviews
Current & peak Elo
Automatic updates · audited input corrections applied. Shared outside-MMA entry and return adjustment; no promotion-specific adjustment. Completed saved UFC events update after the data refresh. Outside history extends when professional MMA eligibility is verified. Calibration refits from earlier results only. Unknown title status receives no bonus; outside career coverage and historical result verification remain incomplete.
Start 1,000 · K=40 · title ×1.10 · KO/TKO or submission ×1.10 · draws 0.5 · no contests excluded. Use the roster selector to include retired fighters. No inactivity deductions. Peak includes the initial 1,000 rating. Verified same-night tournament stages update in sequence. Cancelled, overturned, duplicate, amateur, exhibition and verified non-MMA bouts are excluded.