The paper, the predictions, and a test we can't take back.
Our whitepaper describes the model at the heart of Oraca — a foundation model that reads a player's career like a sequence and forecasts what he becomes — and evaluates it the hard way: trained only on data visible at a historical cutoff, judged on what players actually went on to do.
We publish the back data, not just the results. Every number in the paper can be recomputed from the ledger files on this page. And because backtests are reconstructions, we've also pre-registered a forward test: a frozen set of current predictions, hashed in print, scored publicly from 2027 — whatever the result shows.
Oraca Research · Version 1.0 · 11 June 2026 · DOI 10.5281/zenodo.20652481 · research@oraca.ai
What the paper shows.
Every result below comes from a model that could only see the past, scored against what happened next. The protocol, the leakage controls and the definitions are specified in full in the paper.
37 of 50
Four years of foresight, measured
Given only data visible on 30 June 2021, the model ranked 3,418 under-24 players by their probability of reaching the top 5% of their position by 2025. Its top 50 contained 37 of the 199 who made it — against 32 for the strongest baseline, 31 for ranking by market value, and about three at random.
2 vintages
Not a fluke of one cutoff
A second model, trained independently with a 2017 cutoff and scored against 2020 outcomes, reproduced the result (AUC 0.91). Two clean, leak-controlled backtests, four years apart, both clear of every baseline.
1.95×
The undervaluation screen, nine years running
Across nine consecutive as-of anchors (2016–2024), players flagged by our undervaluation screen moved up a league tier the following season at 1.95 times the base rate — 444 observed where chance predicts 228. Every individual year cleared 1.3×.
§8
What our own validation killed
The paper reports its negative results: an internal estimate that didn't replicate under the stricter protocol, a contaminated backtest we rejected, and where the model's edge narrows. A validation regime that never kills a claim is not a validation regime.
Past model performance is not a guarantee of future results — which is exactly why the forward evaluation below exists.
The pre-registered forward test.
Backtests, however carefully controlled, are reconstructions. So on 11 June 2026 we froze the production model's current predictions — 4,021 players, each with a probability of reaching the top of his position and a probability of drifting out of covered football — and printed the file's cryptographic fingerprint in the paper. The file stays private until scoring, so the predictions can't influence the market they predict. The fingerprint below proves, later, that nothing was changed.
3103caa67e6b64245182778dec31ffd3dbdbb7cfdd6c4c74a2a0e750cd160fcf
11 June 2026
Predictions frozen and hashed
30 June 2027
First public scoring, once the 2026/27 season completes
Annually to 2030
Re-scored and published each season — win or lose
The predictions ledger.
Don't take the paper's word for it. These files contain every prediction behind every published table, joined to what actually happened — enough to recompute the headline numbers yourself in an afternoon. Model outputs released under CC BY 4.0, and permanently archived on Zenodo: DOI 10.5281/zenodo.20652481.
The whitepaper (v1.0)
The full paper: model, protocol, results, negative results, limitations and definitions.
2021 backtest predictions
All 3,418 players scored at the 2021 cutoff, with model rank, probabilities and realized 2025 outcomes.
2017 vintage predictions
The independently trained second vintage: 3,406 players scored at the 2017 cutoff against 2020 outcomes.
Undervaluation screen, nine anchors
49,523 player-seasons (2016–2024): selection flag and realized move-up outcomes per anchor.
Coverage by competition
The corpus behind the model: 34 competitions, first covered season, appearance and player counts.
Ledger README & schemas
Column-by-column schemas, consistency checks and licensing for the files above.
Common questions.
Has Oraca's AI model been backtested?
Yes — on two independent, leak-controlled vintages. A model trained only on data visible at 30 June 2021 was scored against what 3,418 young players went on to do by 2025: its top 50 contained 37 of the 199 who reached the top 5% of their position, versus 32 for the strongest baseline and about three at random. A separately trained 2017-cutoff model reproduced the result against 2020 outcomes.
Can I verify Oraca's published results myself?
Yes. The predictions ledger on this page contains every prediction behind every table in the paper, joined to realized outcomes. Precision, AUC, lift and calibration can all be recomputed from the CSVs; the README documents each column.
What is the pre-registered forward evaluation?
On 11 June 2026 Oraca froze its production model's current predictions for 4,021 players and printed the file's SHA-256 hash in the whitepaper. The file is held privately and will be scored publicly once the 2026/27 season completes (30 June 2027), then annually through 2030 — with results published whatever they show. The hash proves the predictions were never changed.
How does the career forecasting model work?
It treats a player's career as a sequence of monthly observations — minutes, production, club and competition context, valuation, development signals — and trains a generative transformer to predict each next month. Sampling continuations yields a distribution over futures: the probability a player reaches an elite level, and the risk he drifts out of professional football.
What are the model's limitations?
The paper names them: discrimination weakens in the lowest-covered league tiers, probabilities transferred across eras run overconfident at the top, the player-level opponent adjustment operates within league, and an earlier internal estimate of the undervaluation signal's discriminative value did not replicate and is not claimed. The limitations section is part of the paper, not an afterthought.
See the model on your own shortlist.
The research is public; the product is in private beta with a small number of clubs. Tell us where you work and we'll be in touch.