Sports intelligence · WNBA · Operational governance

How a WNBA model earns the right to run

Getting a model into a live workflow is an engineering achievement. Giving it authority is a separate governance decision. This public case study shows the evidence ladder between the two.

Sports modellingShadow operationMonitoringModel risk
See the operating decisionView the governance ladder
1,054,433Candidate observations
6Decision horizons
0Live-capital approvals

· Public operating-model edition.

Decision

Live shadow is observation, not authority.

The public-facing decision is deliberately unambiguous: the operational shadow/live workflow is not live-capital approval. Shadow operation lets the team test timeliness, stability, data quality and review processes against real arrivals while keeping exposure at zero.

Observe first

Let the system ingest, score and record the live-shaped stream without giving the output authority over money or customer decisions.

Compare through time

Evaluate the same event at season open, 24 hours, 12 hours, 6 hours, 2 hours and 30 minutes so late information cannot disguise weak earlier behaviour.

Promote explicitly

Every candidate has a named state. Blocked and research remain visible; shadow is monitored; live-capital approval stays at zero until evidence earns a separate decision.

Operating coverage

Scale tests the system around the model.

More than a million candidate observations exercise ingestion, market mapping, horizon alignment, persistence and review. That volume is evidence of operational coverage—not a count of transactions and not a performance result.

Horizontal bar chart of 1,054,433 WNBA candidate observations: 390,010 sportsbook moneyline, 329,556 spread, 329,474 game total and 5,393 prediction-market moneyline.
The three sportsbook families account for most observed candidates; the smaller prediction-market stream remains visible rather than being hidden in the total.
Data behind the candidate-volume chart
Market familyCandidate observationsPublic interpretation
Sportsbook moneyline390,010Largest observed family.
Sportsbook spread329,556Separate market and settlement shape.
Sportsbook game total329,474Separate target and monitoring surface.
Prediction-market moneyline5,393Smaller stream, reported rather than extrapolated.
Total1,054,433Candidate observations, not trades.

Rolling holdouts

Make the past behave like the future

A rolling holdout trains on earlier information and tests on a later season. That preserves the direction of time and makes changes in league conditions visible. The public evaluation covers 249 holdout events in 2023, 262 in 2024 and 311 in 2025.

Rolling holdout and horizon coverage
Holdout seasonEventsDecision snapshots
2023249Season open, 24h, 12h, 6h, 2h, 30m
2024262Season open, 24h, 12h, 6h, 2h, 30m
2025311Season open, 24h, 12h, 6h, 2h, 30m

Governance ladder

Status is a control, not a slide label.

The portfolio state is intentionally mixed: 16 candidates are in shadow, 2 are under stability watch, 6 remain research-only and 12 are blocked. None has live-capital approval. That distribution is healthy when the system is learning; it would be a problem if every experiment were quietly treated as ready.

Governance chart showing 16 shadow candidates, 2 stability-watch, 6 research, 12 blocked and zero live-capital approved, plus six horizons and rolling holdout counts of 249, 262 and 311.
The ladder combines model state, time-horizon monitoring and historical holdouts so promotion cannot rest on one attractive score.
Long description of the governance-status chart
StateCountAuthority
Shadow16Observed in live-shaped operation; no capital authority.
Stability watch2Additional monitoring required before any promotion decision.
Research6Offline investigation only.
Blocked12Explicitly prevented from promotion.
Live-capital approved0No candidate is authorised to control capital.

Leadership takeaway

Build a path to “yes” that protects “not yet”.

  • Separate technical liveness from decision authority. A daemon can be running while the model remains safely unapproved.
  • Monitor the product at the time decisions happen. Six horizons expose whether quality is dependable early or only appears close to start time.
  • Keep blocked work legible. A visible block is an organisational memory; silently deleting failures invites teams to repeat them.
  • Promote a system, not a score. Data freshness, stability, settlement, review and incident handling matter alongside offline modelling.

Limits

What these aggregates cannot show.

Candidate volume demonstrates coverage, not independence or predictive quality. Holdout counts demonstrate evaluation scale, not current performance. Governance counts show the operating state at publication, not a promise of future promotion. No probability, return or suitability claim should be inferred from this page.