It has become clear that the LB is calculated using data that is included in the train set. But what train data does not have is the most recent data on those coins. Starting September 21 to 1st of February there will be plenty of new observations, and mostly important, a new trend (or continuation of current trend in crypto markets). This is important from two perspective:
\nMy prediction? Utilizing the data that was not shared by G-Research will be the key to winning this competition.
", "messages": [ { "id": "1574200", "postDate": "11/07/2021 10:08:38", "content": "It has become clear that the LB is calculated using data that is included in the train set. But what train data does not have is the most recent data on those coins. Starting September 21 to 1st of February there will be plenty of new observations, and mostly important, a new trend (or continuation of current trend in crypto markets). This is important from two perspective:
\nMy prediction? Utilizing the data that was not shared by G-Research will be the key to winning this competition.
", "rawMarkdown": "It has become clear that the LB is calculated using data that is included in the train set. But what train data does not have is the most recent data on those coins. Starting September 21 to 1st of February there will be plenty of new observations, and mostly important, a new trend (or continuation of current trend in crypto markets). This is important from two perspective:\n- First, it gives additional train rows that are more relevant than the previous observations,\n- Second, already mentioned trend. There are already notebooks utilizing purged timeseries (as per Lopez and deployed in pretty much every notebook for Jane Street competition), but what we actually want is to ignore it and submit one model (or ensemble, we have an option to submit 2 notebooks) that will OVERFIT THE TREND ALL THE WAY to February 1st and hope that this trend will stay for at least next 3 months.\n\n\nMy prediction? Utilizing the data that was not shared by G-Research will be the key to winning this competition.", "votes": null }, { "id": "1574302", "postDate": "11/07/2021 12:03:51", "content": "I think this is 100% correct.
\nAnd btw if you seek recent data, check this out
Looks great. But is the source exactly the same? There are differences between excahnges.
", "rawMarkdown": "Looks great. But is the source exactly the same? There are differences between excahnges.", "votes": null }, { "id": "1574790", "postDate": "11/07/2021 21:54:15", "content": "There will be a supplemental_train.csv that will be available with the next 3 months of data. So The edge won't be that much. There is currently some effort to build the data set in case that matters. At least one mentioned by @yamqwe, and mine (see discussion here: https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/285913). I believe it could matter for other reasons (getting more data before the train, using other data for general design…).
", "rawMarkdown": "There will be a supplemental_train.csv that will be available with the next 3 months of data. So The edge won't be that much. There is currently some effort to build the data set in case that matters. At least one mentioned by @yamqwe, and mine (see discussion here: https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/285913). I believe it could matter for other reasons (getting more data before the train, using other data for general design...).", "votes": null } ], "comments": [ { "id": 1574302, "author_name": "", "author_url": "", "post_date": "11/07/2021 12:03:51", "content": "I think this is 100% correct.
\nAnd btw if you seek recent data, check this out
Looks great. But is the source exactly the same? There are differences between excahnges.
", "votes": null, "replies": [] } ] }, { "id": 1574790, "author_name": "lucasmorin", "author_url": "", "post_date": "11/07/2021 21:54:15", "content": "There will be a supplemental_train.csv that will be available with the next 3 months of data. So The edge won't be that much. There is currently some effort to build the data set in case that matters. At least one mentioned by @yamqwe, and mine (see discussion here: https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/285913). I believe it could matter for other reasons (getting more data before the train, using other data for general design…).
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1574200": "It has become clear that the LB is calculated using data that is included in the train set. But what train data does not have is the most recent data on those coins. Starting September 21 to 1st of February there will be plenty of new observations, and mostly important, a new trend (or continuation of current trend in crypto markets). This is important from two perspective:\n- First, it gives additional train rows that are more relevant than the previous observations,\n- Second, already mentioned trend. There are already notebooks utilizing purged timeseries (as per Lopez and deployed in pretty much every notebook for Jane Street competition), but what we actually want is to ignore it and submit one model (or ensemble, we have an option to submit 2 notebooks) that will OVERFIT THE TREND ALL THE WAY to February 1st and hope that this trend will stay for at least next 3 months.\n\n\nMy prediction? Utilizing the data that was not shared by G-Research will be the key to winning this competition.", "1574302": "I think this is 100% correct. \nAnd btw if you seek recent data, [check this out](https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/285726)", "1574624": "Looks great. But is the source exactly the same? There are differences between excahnges.", "1574790": "There will be a supplemental_train.csv that will be available with the next 3 months of data. So The edge won't be that much. There is currently some effort to build the data set in case that matters. At least one mentioned by @yamqwe, and mine (see discussion here: https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/285913). I believe it could matter for other reasons (getting more data before the train, using other data for general design...)." }, "source": "meta" }