New machine learning framework optimizes football team performance across multiple models

Artificial intelligence has come for one of football’s most contested activities: deciding which players are worth signing, and how much they are truly worth. A new study published in the Journal of Big Data presents a machine learning framework that predicts overall player performance ratings with remarkable precision, offering clubs a data-driven alternative to intuition, scouting folklore, and inflated transfer-market guesswork. The research, led by Keshav Kaushik of Sharda University together with collaborators at institutions across India and Italy, demonstrates that a carefully tuned neural network can outperform a field of strong competitors when tasked with estimating a player’s composite quality from dozens of technical, physical, and tactical attributes.

The study draws its data from the FIFA 20 database, a choice the authors are careful to contextualize. Rather than treating the dataset as a source of video-game trivia, the researchers use it as a structured and standardized proxy for real-world football performance characteristics. The overall rating attached to each player in the dataset functions as a summative assessment — a single number that condenses passing accuracy, dribbling, defensive positioning, pace, strength, and numerous other measurable properties into one score. Because the same attributes are tracked consistently for more than seventeen thousand players using identical measurement conventions, the dataset provides a level of cleanliness and comparability that traditional scouting records rarely achieve. In effect, the researchers argue, it behaves like a well-curated player database of the kind professional clubs maintain internally.

At the heart of the framework is a Multi-Layer Perceptron, a foundational neural network architecture in which layers of artificial neurons transform input features through weighted connections and nonlinear activation functions. The proposed model employs two hidden layers, allowing it to capture interactions between player attributes that simpler linear models would miss — for example, the way a player’s effectiveness in a pressing system might depend jointly on stamina, tackle success, and positional awareness rather than on any single attribute. The network learns by adjusting its weights through backpropagation, iteratively minimizing the gap between predicted and actual ratings across training examples until it converges on a mapping from attribute space to overall quality.

The benchmarking exercise was deliberately rigorous. The MLP was evaluated against Optimized Linear Regression, LightGBM, Random Forest Regression, and XGBoost — the latter two being gradient-boosted decision tree ensembles that have become staples of tabular machine learning and are often the strongest off-the-shelf performers on structured data. The results were decisive. The MLP achieved a coefficient of determination, or R² score, of 99.13, meaning it explained more than ninety-nine percent of the variance in player ratings. Its root mean square error came in at 0.6410 and its mean absolute error at 0.4485 — figures indicating that, on average, its predictions deviated from the true rating by less than half a point on the rating scale. Each of the competing models was beaten on these generalization metrics, a notable result given how frequently gradient-boosting methods dominate structured-data competitions.

Beyond a single train-test split, the authors subjected the model to cross-validation, repeatedly partitioning the data so that performance could be assessed across different subsets. The reported average R², together with its standard deviation, provided evidence that the network’s accuracy is not an artifact of one favorable split but a stable property of the model. Stability matters enormously in practical recruitment settings: a prediction tool that fluctuates wildly depending on which players happen to sit in the training sample would be of limited use to an analytics department making decisions worth tens of millions of euros.

The study’s pipeline extends beyond prediction. Before any model was trained, the researchers applied data analytics and principal component analysis, a dimensionality-reduction technique that projects correlated features onto a smaller set of orthogonal components while preserving as much variance as possible. With dozens of overlapping skill attributes — several measures of shooting, multiple dribbling-related statistics, redundant defensive metrics — PCA serves both to compress the feature space and to expose the underlying axes along which players genuinely differ. Complementing this, the team developed AI-based visualization techniques designed to make the model’s outputs interpretable for recruitment staff, supporting what the authors describe as crisis management and recruitment planning in football clubs. The visual layer translates abstract predictions into decision-ready insights, such as identifying undervalued profiles or flagging squad imbalances before they become competitive problems.

The financial motivation behind the work is difficult to overstate. Football constitutes one of the largest sports markets in the world, and player valuations have shown a persistent upward trajectory, with transfer fees for elite talents routinely crossing the hundred-million-euro threshold. In that environment, a systematic error of even a few rating points — corresponding to a misjudgment about whether a midfielder will adapt to a faster league, or whether a striker’s decline is terminal — can translate into enormous financial losses. The authors position their recommendation model as a tool for player selection, efficiency evaluation, and team distribution: deciding not only who to buy, but how a squad should be assembled around complementary attribute profiles.

The research also reflects a broader shift in sports science. What was once the domain of scouts’ notebooks and anecdotal judgment has, over the past decade, become a computational discipline, with clubs building entire analytics departments and leagues standardizing event data collection. Machine learning sits at the center of that transformation because player quality is fundamentally a multivariate, nonlinear phenomenon. A player’s value to a team is not the sum of isolated skills but an emergent property of how those skills combine, and models capable of learning high-dimensional interactions are naturally suited to the problem. The strong performance of the MLP in this study — outpacing both linear and tree-based baselines — suggests that the smooth, continuous structure of player ratings rewards neural networks’ capacity for fine-grained function approximation.

The authors are candid about the framework’s boundaries. The FIFA-derived dataset, however well structured, remains a proxy rather than a direct record of professional match performance, and the model predicts an aggregate rating rather than future career trajectories, injury risk, or adaptation to new tactical systems. Integrating outputs like these with live match data, physical monitoring, and market valuation models would be the natural next step toward a full recruitment decision-support system. Nevertheless, the results make a clear case that AI-based tools can deliver what the authors call precious tips for talent spotting — quantified, reproducible, and fast enough to scan entire player markets in minutes rather than months.

For clubs operating under financial fair-play constraints, for smaller teams unable to compete on scouting budgets, and for analysts seeking defensible justifications for costly decisions, the study offers a template: assemble clean multivariate player data, reduce its dimensionality thoughtfully, benchmark several model families rather than committing to one, and validate stability through cross-validation before deploying predictions into the high-stakes world of the transfer market. In a sport where fortunes turn on half a rating point, a model that errs by less than half of one may prove to be the sharpest signing of the season.