Research methodology

Evidence. Uncertainty. Accountability.

The publication standard matters as much as the prediction. Our record distinguishes a forecast made in advance from an explanation written afterward.

Current status

The original editorial selections are preserved separately. No trained probabilistic model or audited performance history is represented as live. The structured validation, publication ledger and evaluation system form the foundation for future reviewed forecasts.

Source every fact

Measurements must be traceable to a database observation and an approved source with storage and display rights. Missing values remain UNKNOWN. Narrative generation must not invent statistics; schema validation cannot prove prose is factual, so human source review is mandatory.

Probability and confidence

Win probabilities sum to 100%, conditional on a decisive result. A fighter’s KO/TKO, submission and decision probabilities sum to that fighter’s win probability. Confidence from 0–10 describes the assessment and is distinct from win probability. No model is entitled to claim certainty from incomplete evidence.

Publish before the event

The database timestamps each publication and immediately locks that version. Changes require another reviewed version before the event deadline. Version 1 remains the primary evaluation record. Events and matchup identities referenced by publications cannot be quietly changed.

Evaluate without removing losses

Public evaluation reports accuracy, Brier score and calibration, including sample sizes. Draws, no contests, cancellations and pending results are disclosed separately. Corrections to official results append a new record and retain earlier results.

Limits of the ledger

Application roles cannot overwrite or delete published history. A database owner can change database rules, so independent pre-event exports and backups are required before a production launch can claim externally verifiable timestamps.

Model release requirements

A model needs a registered artifact, training cutoff, methodology and validation summary. Temporal holdout testing, calibration assessment and source licensing review are launch requirements. No outcome is guaranteed.

  1. 01Validate sourced data
  2. 02Construct fighter features
  3. 03Analyze the matchup
  4. 04Estimate probabilities
  5. 05Assess confidence
  6. 06Write the analysis
  7. 07Human review
  8. 08Publish immutable version
  9. 09Record result
  10. 10Evaluate the original
Inspect public evaluation