Combining scout reports with data using large language models
Many clubs are now looking in to large language models, like ChatGPT, to help organise their databases of scouting reports. We have gone one step further and created Twelve GPT Live Scout, a tool for summarising what is known about a player in terms of both previous scout reports (written by humans) and data. The aim is to suggest what scouts should look for in the future.
Here is how it works. First the scouts go out and do their job: writing player reports. One of Twelve’s data analysts has written scouting reports on Florian Wirtz, Jamal Musiala and Xavi Simons from their matches in the Bundesliga and recent Euros. You can download these here. Here as an example of part of these reports on Musiala.
We then create a full Twelve GPT data summary of Musiala. The full report can be downloaded here and the summary slide is shown below.
What Twelve GPT Live Scout does is compare, step-by-step, the scouts conclusions to its own statistical analysis. The aim is to help the next scout who watches Musiala focus on the right things. Here are some examples:
Under the ‘Involvement’ headline, Twelve GPT has picked up on the movements and combinations mentioned in the scout reports. It also makes a link between central areas and half-spaces, identified as important in the data, and how the scout describes Musiala’s positioning. Under ‘Run quality’, Twelve GPT is now asking the scouts to pay more attention to wide areas. This is because the data report identifies wide runs as a weakness in his game.
A strength of this approach is identifying missing pieces of the scouting puzzle. When there aren’t so many observations on a particular aspect of a player’s game, then Twelve GPT asks the scouts to think about what to look for. In this case, there wasn’t so much information on passing and providing for teammates, so Twelve GPT links the data back to potential observations.
These suggestions offer focus on player’s weaknesses, so wide areas feature again, along with crossing of the ball, a skill in which he is weaker. The point of these suggestions is to help the scout focus on certain aspects in the next game.
We are currently applying these methods in a wide variety of ways, including,
Aiding boardroom decision-making. Twelve GPT can help combine the data and the scout reports to give a complete summary of a player.
Preparing scouts for their next assignment. Twelve GPT helps identify strengths and weaknesses of a player and suggests what the scout should look for.
Identifying team fit based on scout reports. We can use our model of how a player fits into a team in order to suggest how well the scouts observations align with a team’s needs.
In fact, the possibilities for this approach are endless. We are already using this with Premier League clubs, but the beauty is that the method is scalable to any club looking to make the vital link between the way scouts describe players and the data-driven approach. Contact us today to see how you can use it too.
