Case study
The interesting part was not producing probabilities. It was building a forecast that could still be inspected after every one of the 104 outcomes was known.
Tournament forecasting is easy to overstate when future results leak into evaluation, missing matches quietly disappear, or simulation choices cannot be reproduced. Presaira was structured so those failure modes stay visible rather than being polished away.
Role & scope
The public project spans the Dixon-Coles and Elo model blend, full-tournament simulation, reproducible prediction artifacts, backtesting, calibration, tests and the product surface that presents the forecast.
Approach
How the system earns the result.
01 — Model
Blend scoring dynamics with team strength.
A Dixon-Coles goal model is combined with Elo ratings so the forecast has both match-score structure and an independent strength signal.
02 — Simulate
Run the tournament, not just isolated fixtures.
Each published run simulates the full competition 50,000 times at a fixed seed, preserving group rules, knockout progression and tournament-level uncertainty.
03 — Guard
Make leakage and ordering errors fail loudly.
Pre-cutoff assertions, non-vacuous negative tests and tie-break ordering tests are part of the evidence. A failed convergence target is committed rather than hidden, with production held at the higher simulation count.
04 — Evaluate
Score the forecast after reality happens.
All 104 World Cup matches are covered and scored, with calibration artifacts, historical backtests and a postmortem that discusses the direction of the model's misses.
Evidence
What can actually be checked.
Tournament coverage
104 / 104
Every 2026 World Cup match is represented in the committed evaluation coverage; none is silently imputed away.
Simulation
50,000×
Full-tournament Monte Carlo iterations per production forecast run, using a fixed seed and versioned provenance.
Evaluation posture
Post-event
Calibration, backtests, incident notes and a tournament postmortem remain inspectable after the event finished.
Limits & boundaries
What the case study does not pretend.
The public repository is a curated extract; the live ingestion, database and publication pipeline are deliberately not published.
A committed convergence check records a failed 0.5 percentage-point target between 10k and 50k simulations instead of rewriting the result as a pass.
The portfolio treatment avoids live-odds and sportsbook language because this is model-evaluation proof, not betting theater.
Publication boundary
Public case-study claims are limited to the repository, committed evaluation artifacts and the live Presaira product. Private operating code and planning material remain outside the website narrative.
From proof to useful work
Where this project maps to real service work.
These links come from the governed project/service evidence map. They are not generic cross-sells and do not widen the claims made above.
Next project · Governed AI system
OpportunityOS