In invasion sports like soccer, basketball, and hockey, the most important actions often happen away from the ball. A striker’s decoy run might pull two defenders out of position, creating a lane for a teammate, yet traditional "box score" statistics—goals, assists, or rebounds—record nothing for that striker. This creates a "valuation gap" where players who provide high strategic utility through movement alone are consistently undervalued.
Recent research has shifted from analyzing discrete event logs (which only track the player in possession) to spatiotemporal tracking data, which monitors the $(x, y)$ coordinates of all 22 players and the ball simultaneously. By combining computational geometry with deep learning, researchers can finally quantify the "invisible" contributions that define elite performance.
The Event-Log Blind Spot
For decades, sports analytics relied on event-based datasets. These logs are "ball-centric": they record a pass at time $t$, a tackle at $t+1$, and a shot at $t+2$. However, as Luca Pappalardo and Paolo Cintia argue in PlayeRank, these "mono-dimensional" stats fail to capture the continuous tactical landscape.
When a model only knows where the ball is, it is blind to the defensive pressure or the passing lanes available to the player. Charbel Merhej and colleagues note in their work on defensive valuation that valuing a defender whose job is to ensure nothing happens is impossible without knowing their positioning relative to attackers. To solve this, the field has moved toward integrating full-pitch coordinates to calculate Pitch Control.
Pitch Control and Dominant Regions
The fundamental intuition behind modern spatial modeling is that the pitch is a resource to be controlled. Joël Estephan, Joachim Gudmundsson, and Michael Horton pioneered the use of Dominant Regions—the area of the field a player can reach before any opponent—to classify the quality of passes (Classification of Passes in Football Matches).
If a player makes a pass into a region heavily dominated by the opponent, the risk is high. Conversely, a player who consistently moves into "open" dominant regions is creating high-value outlets for their teammates. This geometric approach has been adapted by Shunsuke Iwashita to evaluate space creation in Ultimate Frisbee, proving that the physics of "who gets there first" is a universal metric for tactical advantage.
Action Valuation through Deep Learning
Once we have the spatial context, the next challenge is assigning a specific point value to every movement. This is known as Action Valuation (AV).
In DeepHoops, Anthony Sicilia and Konstantinos Pelechrinis use LSTMs to process NBA tracking data, estimating the "Expected Points" of a possession at every millisecond. This allows them to value "micro-actions" like off-ball screens. Similarly, Chen Yanai and the Q-Ball team (Modeling Basketball Games Using Deep Reinforcement Learning) use an Actor-Critic architecture to evaluate real-time decisions, identifying players whose movement manipulates the defense even when they never touch the ball.
In soccer, Tom Decroos and Jesse Davis developed STARSS (A Spatio-Temporal Action Rating System), which uses possession sequences to quantify how much a single dribble or tackle increases the probability of a goal. By comparing current sequences against millions of historical examples, the model can reward the "pivotal" players—like defensive midfielders—who initiate high-value transitions (Jiang and Cai).
Counterfactuals and the "Perfect Moment"
The most recent frontier in this research is Counterfactual Reasoning: asking "what if" a player had moved differently.
Shunsuke Iwashita and Keisuke Fujii introduced VTCS (Valuing Timing by Counterfactual Scenarios) to solve the problem of movement initiation (Evaluating Movement Initiation Timing). By computationally shifting a player's run slightly earlier or later in time and re-calculating the resulting pitch control, they can objectively score whether a player chose the "perfect moment" to break into space.
This leads to a data-driven definition of Creativity. In the un-xPass framework, Jesse Davis defines a creative action as one that is both original (low probability of being chosen by an average player) and useful (significantly increases scoring probability). By modeling the full environment, we can finally see the "hidden" passing lanes that only elite playmakers perceive.
Summary of the Shift
| Feature | Traditional Analytics | Spatiotemporal Analytics |
|---|---|---|
| Data Source | Discrete Event Logs (Pass, Shot) | Continuous Tracking ($x, y$ coordinates) |
| Focus | On-ball player | All 22 players + Ball |
| Metric | Frequency-based (Accuracy, Totals) | Value-based (Pitch Control, Expected Points) |
| Key Insight | What happened? | Why did it happen, and what else was possible? |
By treating the pitch as a dynamic field of influence rather than a series of isolated events, researchers are bridging the gap between raw data and the "gut instinct" of professional scouts, making the invisible tactical depth of sports visible and measurable.