Hi Nate, do you plan to publish each country's stage of elimination probabilities on a daily basis during the World Cup? We have developed a game (SoccerShares.nl) that builds on such probabilities and I would much rather use PELE results as input instead of bookmaker quotes.
1) I believe football manager data is a better source of publicly available player ratings.
2) Transfermarkt does not basing its ratings on player ability alone its based on a player's potential. Footballers peak at about 30, transfermarkt values peak at about 21, if Max Dowman replaced Marcus Rashford in the England squad this would be a massive downgrade but not according to Transfermarkt values.
3) You don't take into account heat, this will be a huge factor at this tournament.
All that said, very impressive model, I imagine it performs so much better than the last soccer model. Certainly the poisson replacement makes so much sense. What would be interesting is if we could use the same method to discount/markup expected goals so that we have a more accurate 'dominance?' heuristic to analyse a specific match.
This is awesome. I love the detail methodology here! I would really like to see the math too. I want to see the formulas to calculating the k values, and various parameters like HFA. How do you mathematically determine HFA in elo points?
In an era of everything LLMs/AI/ML, God is it nice to read through methodology that is, at its core, theory motivating layers of "classical" statistical methods. Love it.
Though I gotta admit, I raised an eyebrow at the part about using Claude to impute the values for missing GDP. Does this mean having Claude write the imputation code, or just feeding Claude the GDP data and having it spit out a spreadsheet? Because the former seems reasonable, while the latter feels shaky and non-deterministic. How do you replicate that in the future when Anthropic retires 4.6?
it seems there's no prior for host-continent advantage -- the pattern where European teams win World Cups in Europe and teams from the Americas win them in the Americas.
In the whole history only two winners crossed over: Brazil in 1958, Germany in 2014. Geography does enter PELE, but I think not for host-continent advantage.
And this matters because perhaps teams from the Americas will have a minor edge this time around?
Hi Nate, do you plan to publish each country's stage of elimination probabilities on a daily basis during the World Cup? We have developed a game (SoccerShares.nl) that builds on such probabilities and I would much rather use PELE results as input instead of bookmaker quotes.
1) I believe football manager data is a better source of publicly available player ratings.
2) Transfermarkt does not basing its ratings on player ability alone its based on a player's potential. Footballers peak at about 30, transfermarkt values peak at about 21, if Max Dowman replaced Marcus Rashford in the England squad this would be a massive downgrade but not according to Transfermarkt values.
3) You don't take into account heat, this will be a huge factor at this tournament.
All that said, very impressive model, I imagine it performs so much better than the last soccer model. Certainly the poisson replacement makes so much sense. What would be interesting is if we could use the same method to discount/markup expected goals so that we have a more accurate 'dominance?' heuristic to analyse a specific match.
This is awesome. I love the detail methodology here! I would really like to see the math too. I want to see the formulas to calculating the k values, and various parameters like HFA. How do you mathematically determine HFA in elo points?
In an era of everything LLMs/AI/ML, God is it nice to read through methodology that is, at its core, theory motivating layers of "classical" statistical methods. Love it.
Though I gotta admit, I raised an eyebrow at the part about using Claude to impute the values for missing GDP. Does this mean having Claude write the imputation code, or just feeding Claude the GDP data and having it spit out a spreadsheet? Because the former seems reasonable, while the latter feels shaky and non-deterministic. How do you replicate that in the future when Anthropic retires 4.6?
Do you plan in the future to use transfermarkt injury data to automatically update your rosters ? I remember your previous models did that.
it seems there's no prior for host-continent advantage -- the pattern where European teams win World Cups in Europe and teams from the Americas win them in the Americas.
In the whole history only two winners crossed over: Brazil in 1958, Germany in 2014. Geography does enter PELE, but I think not for host-continent advantage.
And this matters because perhaps teams from the Americas will have a minor edge this time around?