Model Accuracy
Every prediction backtested against real outcomes.
Nothing enters production until it proves it works on historical data it never saw during training. We publish our accuracy metrics because the numbers are the reason clubs trust us with EUR 10,000 decisions.
Accuracy in your league
See how reliable the model is in your specific market, validated league by league.
Search your leagueHow we validate
Every figure is out-of-sample, scored against real transfer outcomes the model never saw.
See the methodOur data foundation
A proprietary blend of independent sources, cross-validated and normalised across leagues.
What we're built onPosition Shortlist, Squad Analytics, Player Deep Dive
xV Valuation Model
89.4% directional accuracy for U23 players
Our proprietary xV model forecasts a player's market value two years into the future. It powers the ranked shortlists, squad valuations, and deep-dive projections in our reports.
How we validate: Every prediction is tested against real transfer outcomes the model never saw during training. Accuracy is verified per league - we include the per-league breakdown in every report so you can see exactly how reliable the predictions are for your target markets.
Player Deep Dive
Anti-Failure Model
80% precision at identifying signings that won't work out
We define "failure" as a player who, within two years of transfer, accumulated fewer than 30% of available league minutes and saw no market-value increase. The model evaluates maturity signals, availability history, positional risk, and destination club stability to flag high-risk signings before they happen.
How we validate: At the default threshold (P(fail) ≥ 30%), the model flags transfers with 78% precision and 55% recall - it catches roughly half of eventual failures, and 4 in 5 flags are correct. At higher confidence thresholds, precision reaches 97% but recall drops to 12.5%.
Injury Prediction
Injury Prediction Model
7 in 10 high-risk flagged players suffer significant absence
A multi-tier model covering player profile, schedule context, and rolling workload patterns. Calibrated per league because our research showed injury patterns are league-specific - a finding most generic models miss.
How we validate: Models are trained locally per league. The Premier League model is different from the Danish Superliga model. Each achieves higher accuracy than a single global model - validated against real injury data across multiple seasons.
The evidence
The accuracy, visualized.
Every model validated against real outcomes the models never saw during training. No averages hiding the weak spots, every confidence interval on display.
Calibration
Predicted vs realized 2-year growth
Each point is one of 21 leagues. The diagonal is a perfect forecast. Points above the line mean the model under-predicted growth - this is by design. We calibrate conservatively because in scouting, overpaying for an overestimated player costs more than missing upside. Across every market, predicted growth tracks realized growth, never scattered randomly.
Lift over baselines
The model vs. the obvious alternatives
Directional accuracy for U23 players, measured the same way for every method on the same out-of-sample transfers. A random guess across three outcomes (up, down, same) gives 33%. Following the market-value trend gets you to 62%. The model reaches 89.4%.
Where it's strongest
Accuracy by segment, with confidence intervals
We don't average away the weak spots. The model is sharpest on young players (89.4%) and softer on the 23–30 band (78.4%) - older players have less predictable trajectories due to career plateaus, positional shifts, and injury accumulation, not lack of data. The whiskers are verified 95% confidence intervals.
Per-league, no cherry-picking
Directional accuracy across 21 leagues
Every league we cover, tested independently on transfers the model never saw. The whiskers are 95% Wilson confidence intervals - wider where fewer transfers were available. Accuracy holds in the high-80s across the board, not just in the leagues with the most data.
Anti-Failure Screening
Precision vs. recall at rising confidence thresholds
We define "failure" as a player who, within two years of transfer, accumulated fewer than 30% of available league minutes and saw no market-value increase - they never established themselves. The base failure rate is 16.2%. As confidence rises, precision climbs from 78% to 97%, but recall drops - at the highest tier the model catches only 12.5% of failures. At the default threshold (P≥30%), it flags 55% of eventual failures with 78% precision.
Anti-Failure Screening
Transfer failure rate by age band
Failure = fewer than 30% of available minutes and no market-value increase within two years. Under-21 players fail at twice the rate of those signed at 24+. This is the core risk-reward tradeoff in moneyball strategies: the youngest players offer the highest upside and the highest failure risk.
Injury Prediction
Precision improves with each feature tier
Each tier adds measurable accuracy. The base rate is 35%. Adding biographical features reaches 52%. Schedule context pushes to 62%. The full workload model reaches 70% - twice the base rate.
Three-Axis Prediction
Model vs. no-change baseline on valuation level
Assuming 'value stays the same' only gets you R² 0.53. The model reaches 0.97 - closing nearly all of the remaining gap to a perfect score. Validated on 157,359 player-snapshots the model never saw during training.
Three-Axis Prediction · By Age
Valuation R² by age band - model vs. naive
Predicting all ages is the easier problem - older players' values are more stable, so even the naive baseline scores well. Under-24 is where it gets hard: values are volatile, development curves diverge, and the naive baseline falls behind. The model closes the gap across all age bands.
Three-Axis Prediction · Performance Δ
U24 performance delta: R² 0.86 - the hardest prediction
Predicting how a young player's on-pitch output will change is far harder than predicting value. The naive baseline (assume no change) scores R² ≈ 0. For U24 players - where development is most volatile - the model reaches R² 0.86. This is where specialization matters.
Data foundation
Built on a proprietary aggregation of independent sources.
No single data source is sufficient. Our models are trained on a proprietary pipeline that fuses several independent source categories, from longitudinal market valuations to match-level performance data and official league records. We cross-validate between them and normalise across leagues and time periods. The aggregation logic, feature engineering, and cleaning rules are proprietary.
See where our data comes fromValidation methodology
Validated against observable outcomes, not estimates.
Comparing predictions to other predictions proves nothing. Our models are validated against events that actually happened - transfers that completed, leagues players moved to, minutes they accumulated, and call-ups they received. These are facts, not estimates.
Directional accuracy on completed transfers. When we predict a player's value will rise, did they subsequently transfer at a higher fee, move to a higher-tier league, or see their market standing increase? Measured on 9,000+ real transitions the model never saw.
Career-outcome validation. Did flagged players move to a top-5 league? Become regular starters? Receive senior international call-ups? These binary outcomes are objectively verifiable and independent of any valuation methodology.
Strict temporal separation. Every accuracy figure is out-of-sample: models are trained on data up to a cutoff date and scored only on outcomes that occurred after it. No lookahead, no backfit, no self-referencing.
Per-league confidence intervals. We publish Wilson-score 95% confidence intervals for every league, not just a single pooled average. You see exactly where the model is strong and where sample size limits certainty.
Why we publish accuracy
No black boxes. Every report includes the per-league accuracy breakdown for the models used. You see exactly how reliable the predictions are for your specific markets.
Validated on data the models never saw. We use strict train-test splits. The accuracy numbers above come from holdout sets - real transfers and real outcomes that were hidden during model training.
League-specific, not one-size-fits-all. A model trained on the Premier League doesn't automatically apply to the Danish Superliga. We calibrate per league and show you which leagues our models are strongest in.
Continuous improvement. Models are retrained as new transfer windows complete and new outcome data becomes available. Accuracy improves every window.
Frequently Asked Questions
How accurate are football transfer predictions?+
Can you predict if a football transfer will fail?+
How do football analytics companies test their models?+
What age do football players peak in value?+
What data do football scouting tools use?+
See the accuracy in your report.
Every report includes per-league accuracy evidence. You send a brief, we send a PDF.
Order a Report