Improving Soccer Analytics Through Process-Oriented Creation Metrics

Wharton Moneyball hosts Eric Bradlow and Adi Weiner, joined by PhD candidate Jonathan Pipping, discuss the limitations of traditional soccer analytics. They explain that the industry standard, Expected Goals (xG), ignores dangerous attacking sequences that do not end in a shot, which leads to a distorted view of player and team value. By moving from outcome-based metrics to process-oriented models like xG+, teams can find stable, predictive indicators that competitors often miss. This analysis helps sports executives and data practitioners look past surface-level predictions. The advantage lies in understanding the mechanics of the game to make better transfer market decisions and avoid overpaying for past performance that is unlikely to repeat.

The Hidden Cost of Shot-Only Analytics

Most soccer analytics rely on Expected Goals (xG), which only tracks data when a player takes a shot. Pipping explains that this creates a major blind spot: an attack can be dangerous, forcing defensive adjustments and creating pressure, but if the final pass fails or no shot is taken, the xG value remains zero.

Expected goals, it is not this absolute value, this true value that you can observe. It is not a counting stat. It is not a statistic that you will just see and it is like, ah, that is the expected goals. It is a model based statistic.

-- Jonathan Pipping

By relying on xG, teams ignore the creation phase of the game. Pipping’s xG+ framework attempts to solve this by modeling the probability of a shot occurring based on player positioning and pitch location. This shift moves the focus from finishing, which is often noisy and inconsistent, to opportunity creation, which Pipping notes is a more stable and predictive skill.

Why Black Box Models Fail Under Pressure

A recurring theme is the danger of relying on predictive models without understanding their mechanics. When analysts provide coaches with a black box prediction, they often face skepticism because they cannot explain why the model reached its conclusion. Pipping emphasizes that practitioners must be able to break down these models to show which components, such as defensive positioning or passing lanes, drive the output.

If you cannot explain it, then why am I going to put my reputation on the line by implementing this thing that you say, trust me bro, it is going to work.

-- Jonathan Pipping

This is critical in the transfer market. If a team buys a player based on a high xG performance, they may be paying for a hot streak that is unlikely to continue. By separating performance into xG for finishing and xG+ for creation, teams can distinguish between a player who got lucky and one who consistently creates high-value situations.

The Paradox of the Blown Lead

The team also explored the paradox of the blown lead, noting that in basketball and soccer, win probability models often become overconfident when a team is ahead. They observed that losing teams frequently hold a high win probability, sometimes exceeding 70 percent, yet still collapse. This suggests that current models are poorly calibrated for the volatility of late-game scenarios. The implication is that immediate, visible metrics often mask the reality that games are more unstable than models suggest. Relying on these overconfident estimates leads to poor in-game decisions, where coaches may fail to adjust because the math says they are safe.

Key Action Items

• Decompose your metrics: Stop evaluating players on a single output like goals or total xG. Split performance into creation (xG+) and finishing (xG) to identify which skills are repeatable. (Immediate)
• Prioritize stable indicators: Focus investment on metrics that show high year-over-year correlation, such as xG+. This helps you build a roster of consistent performers rather than one-hit wonders. (12-18 months)
• Audit your black boxes: If you use AI or machine learning models for decision-making, force a decomposition. If you cannot explain which variables are driving a prediction, do not use it for high-stakes personnel decisions. (Immediate)
• Account for the creation gap: When evaluating offensive strategies, stop ignoring attacks that end without a shot. Analyze the configuration of players and space to value the pressure created, even if the final outcome was not a shot on goal. (Next quarter)
• Calibrate for volatility: Recognize that win probability models are often overconfident. Build in a buffer for late-game management; assume that leads are more fragile than the standard model suggests. (Ongoing)
• Shift from intervention to association: Use your models to identify high-value players, but avoid using them to dictate specific player interventions without understanding the causal mechanism behind the data. (Long-term)