Chapter 15: Key Takeaways
Player Performance Metrics
Core Definition
Player Performance Metrics: A structured, multi-dimensional framework for evaluating individual soccer players using position-specific statistical measures, normalized for playing time, contextualized against peer groups, and combined into composite profiles that support recruitment, development, and tactical decisions.
The Seven Key Points
1. Single Metrics Are Dangerous
Goals and assists capture only a fraction of a player's contribution. Comprehensive evaluation requires metrics spanning four dimensions: scoring, creation, possession, and defense. Within each dimension, both volume (how much?) and efficiency (how well?) matter.
2. Position Determines the Metric Set
The metrics that define excellence vary radically by position. Applying the same yardstick to a goalkeeper and a striker produces meaningless comparisons.
3. Per-90 Normalization Is Essential but Imperfect
Dividing raw counts by minutes played and multiplying by 90 enables fair comparison across different playing times, but small-sample sizes require either minimum-minute thresholds or Bayesian shrinkage to avoid misleading rates.
4. Players Follow Predictable Age Curves
Performance follows an inverted-U shape with age, but the peak age differs by position and by the type of metric measured. Age curves are central to transfer valuation and recruitment strategy.
5. Form Is Measurable
Rolling averages and exponentially weighted moving averages quantify recent performance relative to baseline. The Form Index (recent / baseline) detects hot and cold spells, while the coefficient of variation captures consistency.
6. Radar Charts Reveal the Whole Player
Standardizing metrics to z-scores or percentiles and plotting them on radar charts produces intuitive multi-dimensional profiles that bridge analytics departments and coaching staff.
7. Similarity Algorithms Power Recruitment
Cosine similarity, Euclidean distance, and clustering algorithms answer the fundamental scouting question: "Who else plays like this player?"
Position-Specific Metric Tables
Goalkeepers
| Metric | What It Measures | Why It Matters |
|---|---|---|
| PSxG - GA | Goals saved above expected | Best single measure of shot-stopping |
| Save percentage | Basic stopping rate | Simple but ignores shot difficulty |
| Cross claim rate | Command of the box | Measures aerial authority |
| Distribution accuracy | Passing from the back | Critical in possession-based systems |
| Clean sheet % | Matches without conceding | Combines team and individual defense |
Defenders
| Metric | Centre-Backs | Full-Backs |
|---|---|---|
| Aerial duel win rate | Primary | Secondary |
| Tackles + interceptions per 90 | Primary | Primary |
| Progressive passes per 90 | Important | Important |
| Carries into final third per 90 | Secondary | Primary |
| Crosses and xA per 90 | Rare | Primary (modern) |
| Errors leading to shots | Critical (negative) | Critical (negative) |
Midfielders
| Metric | Holding | Box-to-Box | Attacking |
|---|---|---|---|
| Tackles + interceptions per 90 | Primary | Important | Secondary |
| Pass completion % | Primary | Important | Secondary |
| Progressive passes per 90 | Important | Primary | Important |
| Progressive carries per 90 | Secondary | Primary | Important |
| xG + xA per 90 | Rare | Important | Primary |
| Key passes per 90 | Secondary | Important | Primary |
| Pressure success rate | Important | Important | Secondary |
Forwards
| Metric | Striker | Winger |
|---|---|---|
| Non-penalty goals per 90 | Primary | Important |
| npxG per 90 | Primary | Important |
| xG outperformance (npG - npxG) | Important | Secondary |
| Shots per 90 | Primary | Secondary |
| xA per 90 | Secondary | Primary |
| Successful dribbles per 90 | Secondary | Primary |
| Touches in penalty area per 90 | Primary | Important |
| Pressing actions per 90 | System-dependent | System-dependent |
Essential Formulas
-
Per-90 normalization: $$\text{Metric}_{p90} = \frac{\text{Raw Count}}{\text{Minutes Played}} \times 90$$
-
Z-score standardization: $$z_i = \frac{x_i - \mu}{\sigma}$$
-
Bayesian shrinkage: $$\hat{\mu}_{\text{player}} = \frac{n}{n + \kappa} \cdot \bar{x}_{\text{player}} + \frac{\kappa}{n + \kappa} \cdot \bar{x}_{\text{population}}$$
-
Aging curve (quadratic): $$\text{Performance}(a) = \beta_0 + \beta_1 a + \beta_2 a^2, \quad a^* = -\frac{\beta_1}{2\beta_2}$$
-
Exponentially weighted moving average: $$\text{EWMA}_t = \alpha \cdot x_t + (1 - \alpha) \cdot \text{EWMA}_{t-1}$$
-
Form Index: $$\text{Form Index} = \frac{\bar{x}_{\text{recent}}}{\bar{x}_{\text{baseline}}}$$
-
Cosine similarity: $$\text{cos}(\mathbf{a}, \mathbf{b}) = \frac{\mathbf{a} \cdot \mathbf{b}}{\|\mathbf{a}\| \|\mathbf{b}\|}$$
-
Coefficient of variation: $$\text{CV} = \frac{\sigma}{\bar{x}}$$
Key Thresholds and Benchmarks
| Parameter | Recommended Value | Rationale |
|---|---|---|
| Minimum minutes (full season) | 900 | ~10 full matches; stabilizes per-90 rates |
| Minimum minutes (half season) | 450 | ~5 full matches |
| Bayesian shrinkage kappa | 900 | Minutes at which 50% weight goes to player data |
| Radar chart axes | 6-10 | Fewer is unreadable; more is cluttered |
| Z-score outlier threshold | z | |
| Cosine similarity "strong match" | > 0.90 | Indicates highly similar playing profiles |
Common Pitfalls to Avoid
| Pitfall | Better Approach |
|---|---|
| Comparing across positions without filtering | Always define a position-specific peer group first |
| Using raw counts to compare players with different minutes | Use per-90 rates with minimum-minute thresholds |
| Treating per-90 rates from 200 minutes as reliable | Apply Bayesian shrinkage or exclude sub-threshold players |
| Including penalties in striker evaluation | Use non-penalty metrics (npG, npxG) alongside totals |
| Ignoring survivorship bias in age curves | Use within-player delta methods and acknowledge selection effects |
| Assuming all xG outperformance is skill | Require multi-season persistence before labeling it skill |
| Confusing cosine similarity (shape) with Euclidean distance (magnitude) | Choose the metric that matches the question: "same style?" vs. "same level?" |
| Building radar charts with inconsistent axis direction | Orient all axes so that "farther from center" means "better" |
Quick Reference: Aging Curve Peaks
| Position | Physical Peak | Technical Peak | Overall Peak |
|---|---|---|---|
| Goalkeeper | 27-29 | 28-32 | 28-31 |
| Centre-back | 25-27 | 27-30 | 27-30 |
| Full-back | 24-26 | 26-28 | 25-28 |
| Central midfielder | 25-27 | 27-30 | 26-29 |
| Winger | 24-26 | 26-28 | 25-28 |
| Striker | 25-27 | 26-29 | 26-29 |
Self-Check Questions
Before moving to Chapter 16, you should be able to answer:
- Why is a multi-dimensional framework necessary for player evaluation, and what are the four main dimensions?
- What is per-90 normalization, and what problems does it solve? What problems remain?
- How does Bayesian shrinkage differ from a hard minimum-minutes threshold, and when would you prefer each?
- What is PSxG - GA, and why is it a better goalkeeper metric than raw save percentage?
- Describe the general shape of an age-performance curve and explain how peak age varies by position.
- What is survivorship bias, and how does it distort aging curve analysis?
- How do you construct a z-score-based radar chart, and what design principles make radar charts effective?
- Explain the difference between cosine similarity and Euclidean distance for player comparison. When would each be more appropriate?
- What is the coefficient of variation, and how does it measure player consistency?
- Describe how K-Means clustering can identify player archetypes within a position group.
Connections to Other Chapters
- Chapter 3 (Statistics): Z-scores, percentiles, and regression underpin all metrics in this chapter.
- Chapter 7 (xG): Non-penalty xG per 90 is a core forward evaluation metric.
- Chapter 8 (xA): Expected assists drive creation metrics for midfielders and wingers.
- Chapter 9 (Passing Networks): Progressive passes and pass completion derive from passing analysis.
- Chapter 12 (Defensive Metrics): Tackles, interceptions, and pressures are inputs to defender profiles.
- Chapter 16 (Team Analysis): Player-level ratings feed into squad depth and team chemistry models.
Looking Ahead
Chapter 16: Team Performance Analysis will shift the lens from individuals to collectives, building frameworks for team style fingerprinting, expected points models, squad depth analysis, team chemistry quantification, and Monte Carlo season simulations.
Keep this summary card handy as a reference while working through later chapters.