There are some questions that arose from the competition. Now the competition ends and prizes are announced. Perhaps it is time to seek for answers:
\n1. Is Coach Id 37252510 a Code for Interim Coach or Unknown?
\nA value count of coach id reveals that 37252510 has a count of about 4,500, at least about 100 times of any other coach. This id appears across more than 220 leagues, and sometimes both teams of a match are coached by ‘37252510’. So it should not be a single person. The question is whether it is a code for interim coach or unknown. If it is a code for unknown, or at least includes such cases, will it affect what we should do with treatment of missing values in coach id?
2. Is it Really that Rating/5 = Expected Goal?
\nA plot of binned mean goals scored in a match against the ratings appears so. But as far as I know, goal prediction models usually involve logarithm and exponential functions. So are the ratings really specially constructed in that perfectly linear way, or is it just a coincidence that in fact the expected goals form an exponential curve with the ratings, but in certain sections it is close to a straight line?
3. Can the Ratings be Used in Other Settings of Models?
\nWhile the ratings look like prediction of goals scored, curiosityQuotient has shown that rating difference forms a sigmoid function with the probability of home win or away win. Goal prediction approach uses some Poisson models to work out probabilities of all possible score lines and then winning probabilities of each side, while models which directly predict home/draw/away normally involve logistic/sigmoid function. Are both approaches mathematically linked and therefore the same set of ratings can be used in both approaches?
4. Can the Jax framework Applied to the Dataset Here?
\nAlthough octosport.io published a tutorial about it, apparently no one attempted this approach. If I do not understand it wrong, it involves construction of a model in formulae which spell out clearly the relationship between all variables involved and the parameters to be estimated. In this case, is it true that if we do not know the exact formulae about the rating features, it is not applicable here? Or we can skip the rating features and apply the Jax framework on some models spelling out the relationship about other available features?
There are some questions that arose from the competition. Now the competition ends and prizes are announced. Perhaps it is time to seek for answers:
\n1. Is Coach Id 37252510 a Code for Interim Coach or Unknown?
\nA value count of coach id reveals that 37252510 has a count of about 4,500, at least about 100 times of any other coach. This id appears across more than 220 leagues, and sometimes both teams of a match are coached by ‘37252510’. So it should not be a single person. The question is whether it is a code for interim coach or unknown. If it is a code for unknown, or at least includes such cases, will it affect what we should do with treatment of missing values in coach id?
2. Is it Really that Rating/5 = Expected Goal?
\nA plot of binned mean goals scored in a match against the ratings appears so. But as far as I know, goal prediction models usually involve logarithm and exponential functions. So are the ratings really specially constructed in that perfectly linear way, or is it just a coincidence that in fact the expected goals form an exponential curve with the ratings, but in certain sections it is close to a straight line?
3. Can the Ratings be Used in Other Settings of Models?
\nWhile the ratings look like prediction of goals scored, curiosityQuotient has shown that rating difference forms a sigmoid function with the probability of home win or away win. Goal prediction approach uses some Poisson models to work out probabilities of all possible score lines and then winning probabilities of each side, while models which directly predict home/draw/away normally involve logistic/sigmoid function. Are both approaches mathematically linked and therefore the same set of ratings can be used in both approaches?
4. Can the Jax framework Applied to the Dataset Here?
\nAlthough octosport.io published a tutorial about it, apparently no one attempted this approach. If I do not understand it wrong, it involves construction of a model in formulae which spell out clearly the relationship between all variables involved and the parameters to be estimated. In this case, is it true that if we do not know the exact formulae about the rating features, it is not applicable here? Or we can skip the rating features and apply the Jax framework on some models spelling out the relationship about other available features?
Hi there, thanks for asking questions. About coach id, it is actually a bad id set when coach are not available. That was not done on purpose at all.
\nAs you have seen in the notebook we have released rating and football model in general are based on scoring model. Rating like Elo, Dixon and Cole, Massey's, Poisson and so one. The output can always be transformed into expected goals using linear or log linear relationships. So it is expected and do not change in fact how we expected these rating to be used.
\nAbout you third question. They are not always mathematically related but give similar results. That said having a predictive distribution gives move flexibility.
\nYou could not use Jax rating directly her as it needs to be updated after each match. The update needs the outcome which was not available in the test set.
", "rawMarkdown": "Hi there, thanks for asking questions. About coach id, it is actually a bad id set when coach are not available. That was not done on purpose at all.\n\nAs you have seen in the notebook we have released rating and football model in general are based on scoring model. Rating like Elo, Dixon and Cole, Massey's, Poisson and so one. The output can always be transformed into expected goals using linear or log linear relationships. So it is expected and do not change in fact how we expected these rating to be used.\n\nAbout you third question. They are not always mathematically related but give similar results. That said having a predictive distribution gives move flexibility.\n\nYou could not use Jax rating directly her as it needs to be updated after each match. The update needs the outcome which was not available in the test set.", "votes": null }, { "id": "1785124", "postDate": "05/11/2022 19:57:59", "content": "Thanks for quick response. Solving some of my puzzles👍
", "rawMarkdown": "Thanks for quick response. Solving some of my puzzles👍", "votes": null } ], "comments": [ { "id": 1785110, "author_name": "octosportio", "author_url": "", "post_date": "05/11/2022 19:46:21", "content": "Hi there, thanks for asking questions. About coach id, it is actually a bad id set when coach are not available. That was not done on purpose at all.
\nAs you have seen in the notebook we have released rating and football model in general are based on scoring model. Rating like Elo, Dixon and Cole, Massey's, Poisson and so one. The output can always be transformed into expected goals using linear or log linear relationships. So it is expected and do not change in fact how we expected these rating to be used.
\nAbout you third question. They are not always mathematically related but give similar results. That said having a predictive distribution gives move flexibility.
\nYou could not use Jax rating directly her as it needs to be updated after each match. The update needs the outcome which was not available in the test set.
", "votes": null, "replies": [ { "id": 1785124, "author_name": "szeyeung", "author_url": "", "post_date": "05/11/2022 19:57:59", "content": "Thanks for quick response. Solving some of my puzzles👍
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1785011": "There are some questions that arose from the competition. Now the competition ends and prizes are announced. Perhaps it is time to seek for answers:\n\n**1. Is Coach Id 37252510 a Code for Interim Coach or Unknown?**\nA value count of coach id reveals that 37252510 has a count of about 4,500, at least about 100 times of any other coach. This id appears across more than 220 leagues, and sometimes both teams of a match are coached by ‘37252510’. So it should not be a single person. The question is whether it is a code for interim coach or unknown. If it is a code for unknown, or at least includes such cases, will it affect what we should do with treatment of missing values in coach id? \n\n**2. Is it Really that Rating/5 = Expected Goal?**\nA [plot](https://www.kaggle.com/code/szeyeung/ratings-are-prediction-of-goals-scored****) of binned mean goals scored in a match against the ratings appears so. But as far as I know, goal prediction models usually involve logarithm and exponential functions. So are the ratings really specially constructed in that perfectly linear way, or is it just a coincidence that in fact the expected goals form an exponential curve with the ratings, but in certain sections it is close to a straight line?\n\n**3. Can the Ratings be Used in Other Settings of Models?**\nWhile the ratings look like prediction of goals scored, [curiosityQuotient ](https://www.kaggle.com/code/curiosityquotient/investigation-into-rating-feature-octosport)has shown that rating difference forms a sigmoid function with the probability of home win or away win. Goal prediction approach uses some Poisson models to work out probabilities of all possible score lines and then winning probabilities of each side, while models which directly predict home/draw/away normally involve logistic/sigmoid function. Are both approaches mathematically linked and therefore the same set of ratings can be used in both approaches?\n\n**4. Can the Jax framework Applied to the Dataset Here?**\nAlthough octosport.io published a [tutorial](https://www.kaggle.com/code/octosportio/rating-and-predicting-e-sports-with-jax/comments) about it, apparently no one attempted this approach. If I do not understand it wrong, it involves construction of a model in formulae which spell out clearly the relationship between all variables involved and the parameters to be estimated. In this case, is it true that if we do not know the exact formulae about the rating features, it is not applicable here? Or we can skip the rating features and apply the Jax framework on some models spelling out the relationship about other available features?", "1785110": "Hi there, thanks for asking questions. About coach id, it is actually a bad id set when coach are not available. That was not done on purpose at all.\n\nAs you have seen in the notebook we have released rating and football model in general are based on scoring model. Rating like Elo, Dixon and Cole, Massey's, Poisson and so one. The output can always be transformed into expected goals using linear or log linear relationships. So it is expected and do not change in fact how we expected these rating to be used.\n\nAbout you third question. They are not always mathematically related but give similar results. That said having a predictive distribution gives move flexibility.\n\nYou could not use Jax rating directly her as it needs to be updated after each match. The update needs the outcome which was not available in the test set.", "1785124": "Thanks for quick response. Solving some of my puzzles👍" }, "source": "meta" }