## The Quarterly Model Review
Expert-level model management includes a formal quarterly review that evaluates every component of model performance and generates specific improvement actions.
## The Review Checklist
**Performance metrics:**
- Log-loss (rolling 500 matches): above or below baseline?
- Average CLV (rolling 500 bets): positive, neutral, or negative?
- Calibration curve: any systematic bias at specific probability ranges?
- ROI by market type: which markets are profitable, which are not?
**Model health checks:**
- Feature importance: are the expected features most predictive? (If unexpected features dominate, investigate)
- Coefficient stability: have model coefficients changed significantly from last quarter? (Large changes suggest data drift)
- Residual distribution: are residuals normally distributed? Systematic patterns indicate model misspecification
**Data quality:**
- Missing data rate: higher than last quarter? (Indicates data source issues)
- Feature distribution: significant shifts vs last quarter? (Indicates potential concept drift)
- Match coverage: are all target leagues fully covered?
**Market comparison:**
- Opening vs closing CLV split: is edge concentrated at opening or maintained through closing? (Opening concentration suggests information is quickly priced in — threat to edge longevity)
- CLV by bookmaker: which bookmakers offer the best lines? Have any changed?
## The Improvement Backlog
Each review produces a prioritised improvement backlog:
- High priority: features that should improve calibration based on clear error pattern
- Medium priority: additional data sources to integrate
- Low priority: algorithmic improvements to investigate in the next research cycle
## The Long-Run Perspective
A model review protocol applied consistently over 3+ years produces a richly documented model evolution history: what was changed, why, and what effect it had. This history is the intellectual property of the operation — irreplaceable and compounding in value.
Create a free account to track your progress and save bookmarks.
