Career Outlook

Most scouting data tells you how good a player is today. Career Outlook answers the question that actually decides a transfer: what does this player become? It is a career foundation model — the same family of AI that powers modern language models — trained on two decades of complete professional careers, month by month: minutes, production, the strength of the club and league, market value, injuries, moves. Ask it about a young player and it doesn't offer an opinion: it simulates hundreds of plausible futures for him and reports the probability distribution over how they end — the chance he reaches the elite tier of his position, the path his level is likely to take, and the risk his career stalls.

In one line: a language model learns what tends to come next in a sentence by reading millions of them; Career Outlook learns what tends to come next in a career by reading tens of thousands of them. It plays a player's career forward hundreds of times and counts the endings — and the count is the probability.

Careers, read like language

A professional career is a sequence with its own grammar: the breakthrough, the plateau, the move up that comes a season too early, the injury that resets a trajectory, the quiet year at a smaller club that precedes the loud one. The model reads each career as exactly that — a month-by-month sequence in which every step records how much a player played, what he produced, how strong his club and league were, what the market thought of him, whether he was fit, and whether he moved — and learns, across two decades of careers, how such sequences tend to continue.

Even what's missing is treated as signal rather than noise: which data exists for a player is itself information about where and when he has been playing, so gaps are read, not papered over.

To assess a player, the model then writes his future the way a language model writes text: one month at a time, hundreds of separate times. Some of those simulated futures see him reach the top of his position; some see him plateau; some see him drift out of covered football altogether. The share of futures in which he reaches the top 5% of his position — on the same cross-league ability scale used throughout Oraca — within four years is the headline probability. And because whole careers are simulated, the same futures also carry how much he is likely to play, the level he is likely to settle at, and the other tail: the chance the career goes nowhere, which protects a budget as surely as finding a star does.

Career Outlook is the newest member of the projection family: where Trajectory & Ceiling Forecasting projects a compact summary of a career forward, Career Outlook generates the careers themselves and reads every answer off the same set of simulations.

Tested the only way that counts: against the future

We don't grade our models on the past they trained on. We rewound to July 2021 and rebuilt everything as if the years since had never happened — the model, and every choice around it, saw nothing after that date. Then we asked it to rank all 3,422 under-23 players visible in our covered leagues at the time, and waited four years.

• Of its top 50 picks, 37 became top-5% players at their position by 2025 — versus roughly 3 by chance, about 23 from ranking the same players by current ability, and about 31 from ranking them by market value.
• Its 2021 top 25 included Pedri at #1, Mbappé, Saka, Foden, Ødegaard and Rice — alongside Haaland, whom the model had been reading since his Molde months, and Gravenberch, then playing for Jong Ajax. Twenty of the twenty-five made it, and the five that didn't all finished inside the top 8% of their positions.
• We then repeated the entire experiment from 2017 with a separately trained model, so the result wouldn't rest on one lucky era. Same story: its shortlist was about 11× better than chance.
• Every published probability is calibrated: when the model says 70%, it happens about 70% of the time. A trustworthy 25% is worth more to a recruitment meeting than a confident-sounding guess.

Past model performance is not a guarantee of future results — which is why the backtest is only the beginning.

A track record logged before the fact

However disciplined, a backtest is still a reconstruction. So from June 2026, every monthly Career Outlook prediction is logged before the fact and graded as the careers actually unfold. Once a month's predictions are published they are never revised — the log only grows. Over time that becomes the strongest evidence a forecasting system can offer: a track record built in the open, not a story told in hindsight.

For every level of the game

"Top 5% of his position" is a headline, not the limit. Because the model simulates whole careers, the same futures answer different questions for different clubs: a top-flight director sees the players most likely to reach the world's elite; a League One sporting director also sees the players most likely to become his level's difference-makers — and the ones most likely to be playing Championship football in three years. The question changes with the club; the simulated futures underneath are the same.

What it can't do

• Coverage, not model. The model sees 30 leagues deeply. A prospect in an uncovered league is invisible until he arrives in covered football, and recent arrivals carry a limited-history flag while the model learns them.
• It doesn't see this morning's scan. The probability does not embed a player's current injury status, so the product always pairs it with his live injury context — the number without it would mislead.
• Monthly, not minute-by-minute. Outlooks are computed as monthly vintages; a mid-month transfer or setback shows up in the next run, not the current one.
• Distributions, not verdicts. A 25% chance of stardom is a precise statement about an uncertain future, and it is shown as one — never rounded up to "yes" or down to "no".
• No model decides alone. Career Outlook ranks who to watch; your scouts decide who to sign.

The research behind it

Career Outlook applies the modern language-model recipe — and the older science of calibrated probability forecasting — to careers instead of text.

• Vaswani, A. et al. (2017). Attention is all you need. — The neural architecture the career model is built on.
• Radford, A. et al. (2019). Language models are unsupervised multitask learners. — The generative recipe: train on whole sequences, then read answers off what the model writes forward.
• Bommasani, R. et al. (2021). On the opportunities and risks of foundation models. — The foundation-model framing: one model of an entire domain, many questions asked of it.
• Zadrozny, B. & Elkan, C. (2002). Transforming classifier scores into accurate multiclass probability estimates. — The calibration step that makes a published 70% mean 70%.
• Gneiting, T. & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. — The standard for judging probabilistic forecasts honestly: against outcomes, as distributions.