TL;DR
Researchers have developed a predictive model to quantify defensive effectiveness in the Chinese Football Super League (CSL) by linking specific playing styles to "Expected Goals" (xG). The study reveals that while preventing "dangerous situations" is the most critical factor, a "deep-sitting" defense combined with high-intensity confrontation is the most effective way to stifle opponents in the CSL.
What Problem Does This Paper Address?
In football analytics, the spotlight usually shines on the attackers—goals, assists, and creative passing are easy to quantify. Defense, however, is notoriously difficult to measure. It often involves reactive behavior, off-ball positioning, and tactical discipline that doesn't always show up in a simple box score.
The authors address the gap in "prescriptive" defensive modeling. While previous studies have classified how teams defend, few have successfully measured the effectiveness of those styles. By focusing on the Chinese Super League (CSL), the researchers aim to provide coaches with a data-driven roadmap for which defensive behaviors actually lead to fewer high-quality scoring chances for the opposition.
Methods
The researchers employed a three-step statistical approach:
- Principal Component Analysis (PCA): They took 15 defense-related performance indicators (like clearances, keeper claims, and "ball gains" in specific zones) and used PCA to group them into 8 distinct "factors" or defensive styles. This reduced the complexity of the data while retaining over 85% of the variance.
- Expected Goals (xG) Modeling: Using a dataset of nearly 28,000 shots, they built a logistic regression model to calculate the probability of any given shot resulting in a goal based on variables like pitch location, body part used, and the type of preceding action (e.g., a cross vs. a through ball).
- Multiple Linear Regression: Finally, they used the defensive styles (the PCA factors) as independent variables to predict the opponent's total xG in a match. This allowed them to see which styles had a statistically significant impact on reducing the opponent's scoring threat.
Datasets
- Match Data: Event data from all 1,120 matches played in the CSL across five seasons (2016–2020).
- Shot Data: 27,852 individual shots used to train and validate the xG model.
- Source: Data was retrieved from Opta Sports (via WhoScored.com), a standard benchmark for high-fidelity football event data.
Key Insights and Ideas
- The "Dangerous Situation" Trap: The most significant predictor of conceding high xG was Factor 7 ("Receiving a dangerous situation"). This suggests that defensive failure in the CSL is often a result of being slow in transition, allowing opponents to create "imbalanced" situations where the defense is out of position.
- Deep Defense Wins in the CSL: Contrary to the "high-press" trend popular in European leagues (like the Bundesliga), the model found that "Defense closed to the own goal" (Factor 1) was highly effective at reducing opponent xG. In the CSL context, dropping deep to protect the box is a more reliable strategy than pressing high up the pitch.
- The Goalkeeper’s Proactive Role: "Keeper claims" (Factor 8) emerged as a significant negative predictor of opponent xG. This highlights that a goalkeeper's ability to intercept crosses and long balls before they turn into shots is a vital, yet often overlooked, defensive "style."
Key Results
- Model Accuracy: The xG model achieved an AUC (Area Under the Curve) of 79.83%, indicating high reliability in distinguishing between high- and low-quality scoring chances.
- Variance Explained: The final regression model explained 41.6% of the total variance in opponent xG. While this leaves room for other factors (like individual player brilliance), it is a substantial figure for team-level tactical analysis.
- Predictive Power: When tested on the 2020 season data, the model's predicted xG correlated with actual opponent xG at r = 0.62, a moderate-to-strong correlation that validates the model's practical utility for coaches.
- Top Predictors: "Receiving a dangerous situation" accounted for 31.3% of the variance in opponent xG, followed by "Defense closed to the own goal" at 7.6%.
Limitations and Future Work
The authors acknowledge several areas for improvement:
- Missing "Off-Ball" Data: The study relies on event data (on-ball actions). It does not include tracking data, which would capture player positioning and physical metrics (sprinting, distance covered) that are crucial for a complete defensive picture.
- Contextual Nuance: The model doesn't fully account for "match status" (e.g., a team defending a lead vs. chasing a game) or the quality of the opposition, which can drastically alter defensive behavior.
- League Specificity: The findings are tailored to the CSL. Because playing philosophies vary globally, a "deep-sitting" defense might be less effective in a league with higher crossing accuracy or better long-range shooters.
Future research aims to integrate spatio-temporal tracking data to better understand how defensive shapes—not just actions—influence the game.
Summary updated