The Second Coach

Inspiration

Football analytics has become extremely strong at measuring outcomes such as passes completed, shots generated, and expected goals, yet coaching decisions are not made on outcomes alone but on the quality of the choices taken under pressure. In many situations a safe pass will protect a player’s completion rate while quietly reducing attacking potential, whereas a progressive forward pass may fail technically but still represent the correct decision given the context. This gap between outcome evaluation and decision evaluation is the structural flaw we set out to address by focusing on the opportunity cost embedded in every on ball action and transforming it into a measurable and actionable signal.

What it does

The Second Coach evaluates each player action against the realistic alternatives available in that exact game state and quantifies how much value was gained or lost relative to the optimal choice. At the moment of the touch we reconstruct the full spatial and tactical context of the pitch, including pressure, defensive density, passing lanes and match situation, and we model the credible options the player could have executed. For each alternative we estimate the probability that the team will score within the next five actions, and we calculate the Threat Gap between the selected action and the best available option. The result is a Decision Quality score that reflects whether the player increased team value or left opportunity on the table, independent of whether the execution succeeded.

How we built it

We trained an XGBoost model on more than 500,000 events collected across seven competitions and engineered 28 contextual features that capture pitch location, opponent proximity, directional pressure, angle to goal and game state. The system extracts the exact context at the millisecond of the action, generates feasible alternatives constrained by real spatial structure, and predicts the downstream scoring probability associated with each option over the next five actions. By comparing the chosen action with its optimal alternative we isolate decision making from technical execution and create a robust quantitative measure of intelligence under pressure.

Challenges we ran into

One of the main technical challenges was evaluating actions that never occurred, since modeling realistic alternatives requires balancing tactical plausibility with statistical rigor. We had to ensure that the generated options reflected true constraints on the pitch rather than artificial scenarios introduced by the model. Another key challenge was redesigning the evaluation framework so that progressive and creative play is not systematically penalized, which required careful feature validation and bias testing to prevent the model from rewarding conservative possession by default.

Accomplishments that we're proud of

We developed a complete end to end pipeline that transforms raw event data into a clear and interpretable Decision Quality metric, introduced the Threat Gap as a direct measure of opportunity cost, and demonstrated the model on a real Bundesliga match to show how it surfaces hidden inefficiencies and undervalued performance. Beyond the technical achievement, we reframed the analytical standard from evaluating what happened to evaluating what should have happened, creating a foundation for more objective performance feedback, tactical diagnostics and smarter scouting decisions.

What we learned

Through this process we learned that outcome driven metrics can systematically mask poor decision making and that risk aversion is often statistically rewarded in traditional frameworks. When context is properly reconstructed and alternatives are modeled realistically, decision intelligence becomes measurable and comparable across players and roles. Opportunity cost, once quantified, provides a powerful lens to evaluate contribution beyond surface level statistics.

What's next for The Second Coach

The next phase is to integrate richer tracking data to further strengthen contextual reconstruction and alternative modeling, expand the framework into structured player benchmarking and scouting workflows, and explore real time applications that support coaching staff during match analysis. Our ambition is to establish decision intelligence as a core performance layer that improves competitive outcomes while reducing financial risk in player evaluation and recruitment.

Built With

• javascript
• python