Multi-axis projection
Three-Axis Prediction
Value, performance, and minutes - projected two years forward.
A complementary engine that scores every player on three forward-looking axes at once on a two-year horizon, delivering 17x more predictive lift over naive baselines than a short-term model.
What it does
Three-Axis projects each player forward on value, performance, and minutes simultaneously over a two-year horizon, then looks for agreement across the three. A naive no-change baseline achieves R² 0.53 on two-year valuation - the model reaches 0.77, closing almost half the remaining gap. That lift is where the scouting signal lives: identifying the player whose value is about to move before the market reprices them.
Used in: Cross-model shortlisting and breakout detection.
Why it’s defensible
What makes this proprietary.
Measured against what matters: lift over naive
A short-horizon model can report R² 0.96 because values barely change over five months - the no-change baseline already reaches 0.94. Over two years, the no-change baseline drops to 0.53 and the model delivers 17x more lift. Raw accuracy on an easy target is not the same as genuine predictive power on a hard one.
Agreement across independent axes
The power of the method is in the intersection: three separately-trained projections converging on the same player. That convergence is a much higher bar than a single model clearing a threshold, and it is why the combined shortlist outperforms any individual axis.
Leakage-controlled validation
The performance and minutes axes are validated on held-out pairs with explicit controls against information leakage, so the correlation figures reflect genuine forward predictive power rather than an artifact of the test design.
How we validate it
Each axis is validated independently against observable outcomes - valuation direction against completed transfers and league-tier moves, performance and minutes against held-out, leakage-controlled outcome pairs. The headline metric is lift over two naive baselines: no-change (assume current value persists) and cohort-mean (assume every player converges to the average). The model outperforms both substantially on every axis.
What we don’t publish
The three projection models, how their outputs are combined, and the leakage controls are proprietary. We publish the validated accuracy of each axis, the lift over naive baselines, and the confidence intervals - so the result can be trusted without exposing how it is produced.
The underlying data and models are proprietary. We show the validated results, not the inputs that produce them.
See it applied to your shortlist.
Every report shows the model’s output with the per-league accuracy behind it.
Order a Report