I'm happy to announce that we are releasing an interim update to the supplemental_train.csv file. This will enable you to get a head start on updating your models to incorporate data gathered since the competition began. It should also reduce the risk of error from any workarounds to handle the old placeholder data and provides an opportunity for you to test your models on this updated data before submissions close on 1 Feb. None of the other files will change. The update should be live within the next hour.
\nThere will still be one more update to the supplemental_train.csv on the original schedule (after the submission deadline), which will cover the remainder of the period up until the submission deadline. Barring any missing data, the final copy of supplemental_train.csv will end right before the forecasting phase test set data begins.
\nEdit: @tomforbes has identified some implausible values in the updated supplement. We are looking into the matter and will provide an update when we have more information.
\nEdit 2: There was indeed some erroneous data coming from the data provider, which have been corrected and updated in the supplemental data file. The corrections should be available to your notebooks within the next hour.
", "messages": [ { "id": 1650142, "postDate": "2022-01-14T20:45:50.973Z", "content": "I'm happy to announce that we are releasing an interim update to the supplemental_train.csv file. This will enable you to get a head start on updating your models to incorporate data gathered since the competition began. It should also reduce the risk of error from any workarounds to handle the old placeholder data and provides an opportunity for you to test your models on this updated data before submissions close on 1 Feb. None of the other files will change. The update should be live within the next hour.
\nThere will still be one more update to the supplemental_train.csv on the original schedule (after the submission deadline), which will cover the remainder of the period up until the submission deadline. Barring any missing data, the final copy of supplemental_train.csv will end right before the forecasting phase test set data begins.
\nEdit: @tomforbes has identified some implausible values in the updated supplement. We are looking into the matter and will provide an update when we have more information.
\nEdit 2: There was indeed some erroneous data coming from the data provider, which have been corrected and updated in the supplemental data file. The corrections should be available to your notebooks within the next hour.
", "rawMarkdown": "I'm happy to announce that we are releasing an interim update to the **supplemental_train.csv** file. This will enable you to get a head start on updating your models to incorporate data gathered since the competition began. It should also reduce the risk of error from any workarounds to handle the old placeholder data and provides an opportunity for you to test your models on this updated data before submissions close on 1 Feb. None of the other files will change. The update should be live within the next hour.\n\nThere will still be one more update to the **supplemental_train.csv** on the original schedule (after the submission deadline), which will cover the remainder of the period up until the submission deadline. Barring any missing data, the final copy of **supplemental_train.csv** will end right before the forecasting phase test set data begins.\n\nEdit: @tomforbes has identified some implausible values in the updated supplement. We are looking into the matter and will provide an update when we have more information.\n\nEdit 2: There was indeed some erroneous data coming from the data provider, which have been corrected and updated in the supplemental data file. The corrections should be available to your notebooks within the next hour. ", "votes": 43 }, { "id": 1653199, "postDate": "2022-01-17T10:36:42.580Z", "content": "Hi @sohier as mentioned by @tomforbes the new data has rows that are clearly wrong prices. For example, the price of Tron changed in 1-minute bar 6888% several times on October 29. The important fact is that the training set seems to be already curated filtering out any exchange or flash crashes while the supplemental not. These types of returns have a huge impact on modeling as are completely random and have several orders of magnitude. Should we assume that the final data set will include these type of price action?
", "rawMarkdown": "Hi @sohier as mentioned by @tomforbes the new data has rows that are clearly wrong prices. For example, the price of Tron changed in 1-minute bar 6888% several times on October 29. The important fact is that the training set seems to be already curated filtering out any exchange or flash crashes while the supplemental not. These types of returns have a huge impact on modeling as are completely random and have several orders of magnitude. Should we assume that the final data set will include these type of price action? ", "votes": 7, "replies": [ { "id": 1653892, "postDate": "2022-01-18T00:46:40.827Z", "content": "Compared to Binance data, I can confidently say that this data contains errors. Now, what I wonder is if this error will be fixed soon, and the final supplementary data set includes such errors.
", "rawMarkdown": "Compared to Binance data, I can confidently say that this data contains errors. Now, what I wonder is if this error will be fixed soon, and the final supplementary data set includes such errors.", "votes": 3 }, { "id": 1657115, "postDate": "2022-01-19T23:14:01.040Z", "content": "I've just uploaded a correction to the new data; it should be available to your notebooks soon.
", "rawMarkdown": "I've just uploaded a correction to the new data; it should be available to your notebooks soon.", "votes": 5 }, { "id": 1657129, "postDate": "2022-01-19T23:46:10.507Z", "content": "Thanks for the update!
", "rawMarkdown": "Thanks for the update!", "votes": 1 } ] }, { "id": 1651166, "postDate": "2022-01-15T15:39:03.813Z", "content": "Hi @sohier, there appears to be some errors in the update which i have outlined here
", "rawMarkdown": "Hi @sohier, there appears to be some errors in the update [which i have outlined here](https://www.kaggle.com/tomforbes/errors-in-supplemental-train-csv)", "votes": 8, "replies": [ { "id": 1652996, "postDate": "2022-01-17T06:51:42.187Z", "content": "Many thanks for pointing this out, I experienced similar issues. Hopefully the Kaggle staff and organizers can give us some clarity on this.
", "rawMarkdown": "Many thanks for pointing this out, I experienced similar issues. Hopefully the Kaggle staff and organizers can give us some clarity on this." }, { "id": 1654871, "postDate": "2022-01-18T22:28:36.083Z", "content": "Thank you very much, @tomforbes, @joseantonioalatorre and other kagglers! Three days ago you posted this valuable info. Unfortunately, nobody inside Kaggle seems to care..
", "rawMarkdown": "Thank you very much, @tomforbes, @joseantonioalatorre and other kagglers! Three days ago you posted this valuable info. Unfortunately, nobody inside Kaggle seems to care.." }, { "id": 1654876, "postDate": "2022-01-18T22:35:29.970Z", "content": "@tomforbes thank you for flagging this, we are looking into the matter.
", "rawMarkdown": "@tomforbes thank you for flagging this, we are looking into the matter.", "votes": 4 }, { "id": 1654884, "postDate": "2022-01-18T22:57:43.320Z", "content": "Thank you, @sohier and apologies for my previous claim. I see that with the added data, some targets can be updated also.
", "rawMarkdown": "Thank you, @sohier and apologies for my previous claim. I see that with the added data, some targets can be updated also. ", "votes": 1 } ] }, { "id": 1651312, "postDate": "2022-01-15T17:11:11.970Z", "content": "Why is this 'missing one minute' each month? Yes I am talking about the first minute of 4 months in the dataset?
", "rawMarkdown": "Why is this 'missing one minute' each month? Yes I am talking about the first minute of 4 months in the dataset?\n", "votes": 1 }, { "id": 1650899, "postDate": "2022-01-15T12:05:07.770Z", "content": "Is there time interval between training data and test data ?
", "rawMarkdown": "Is there time interval between training data and test data ?", "votes": 1, "replies": [ { "id": 1651322, "postDate": "2022-01-15T17:14:09.927Z", "content": "In the evaluation phase, there will not be any data gap. So you can safely assume that your model will start predicting just after the last minute of data used in training… except the one minute that that is mentioned below.
", "rawMarkdown": "In the evaluation phase, there will not be any data gap. So you can safely assume that your model will start predicting just after the last minute of data used in training... except the one minute that that is mentioned below.", "votes": 2 }, { "id": 1652776, "postDate": "2022-01-17T01:30:45.013Z", "content": "Hello @abdulkadirguner, based on the original train and test datasets, I believe we are predicting for the next 3 months every minute into the future. Do you think this will be the case after the competition is close, trying to understand how much are we predicting / forecasting into the future?
", "rawMarkdown": "Hello @abdulkadirguner, based on the original train and test datasets, I believe we are predicting for the next 3 months every minute into the future. Do you think this will be the case after the competition is close, trying to understand how much are we predicting / forecasting into the future?", "votes": 1 }, { "id": 1652852, "postDate": "2022-01-17T03:55:55.723Z", "content": "@cv12j0, you can say that. Competition ends beginning of February. Submitted notebooks will run against real data in the next 3 months period (until beginning of May). With one nuisance: this evaluation phase is divided into 2-week periods. So every two weeks, your notebooks will be assessed with data from the beginning of the test phase to the end of that two weeks. So if you are end of week 4, you will be assessed for the whole 4 weeks, and at the end of week 6, for the whole 6 weeks. Buttt, there is only one final score and that is your score for the whole 3-months of data, from start to finish at then end of week 12(end of month3).
", "rawMarkdown": "@cv12j0, you can say that. Competition ends beginning of February. Submitted notebooks will run against real data in the next 3 months period (until beginning of May). With one nuisance: this evaluation phase is divided into 2-week periods. So every two weeks, your notebooks will be assessed with data from the beginning of the test phase to the end of that two weeks. So if you are end of week 4, you will be assessed for the whole 4 weeks, and at the end of week 6, for the whole 6 weeks. **Buttt, there is only one final score and that is your score for the whole 3-months of data, from start to finish at then end of week 12(end of month3)**.", "votes": 1 }, { "id": 1652857, "postDate": "2022-01-17T04:08:02.497Z", "content": "Thanks, @abdulkadirguner trying to build lag features that can handle any future condition.
", "rawMarkdown": "Thanks, @abdulkadirguner trying to build lag features that can handle any future condition." }, { "id": 1652864, "postDate": "2022-01-17T04:18:09.900Z", "content": "It seems that the current supplemental_train.csv file is completely different from train.csv. 👀
", "rawMarkdown": "It seems that the current supplemental_train.csv file is completely different from train.csv. 👀", "votes": 1 }, { "id": 1653990, "postDate": "2022-01-18T03:13:05.367Z", "content": "Yes @yuanzhezhou it is, but the main reason is to provide extra data, as they explain in the data description section
", "rawMarkdown": "Yes @yuanzhezhou it is, but the main reason is to provide extra data, as they explain in the data description section" } ] }, { "id": 1654893, "postDate": "2022-01-18T23:18:12.417Z", "content": "@sohier In the forecasting phase, since the supplemental_train.csv is already fixed, there will be time gaps between supplemental_train.csv and test after 1st update, am I understanding correctly?
\nI would also like to know for sure, but I believe the test will be rerun from the same starting point each update like:
\nweek 0-2
\nweek 0-4
\nweek 0-6
\n…..
\nweek 0-12 (final update)
This is a good point actually since either we time the notebook so that it is sufficient just to run for 2 weeks or the full 12 weeks. Feature engineering time will be limited for the latter case.
", "rawMarkdown": "This is a good point actually since either we time the notebook so that it is sufficient just to run for 2 weeks or the full 12 weeks. Feature engineering time will be limited for the latter case." }, { "id": 1656213, "postDate": "2022-01-19T07:14:43.720Z", "content": "week 0-12 will take a lot of time to compute then..
", "rawMarkdown": "week 0-12 will take a lot of time to compute then.." }, { "id": 1656807, "postDate": "2022-01-19T16:32:01.233Z", "content": "Each re-run will start from the same time point, with the final re-run (the only one that matters) including 3 months of data. The Public Leaderboard has similar amount of data and should be used to test your running time.
", "rawMarkdown": "Each re-run will start from the same time point, with the final re-run (the only one that matters) including 3 months of data. The Public Leaderboard has similar amount of data and should be used to test your running time.", "votes": 8 }, { "id": 1657237, "postDate": "2022-01-20T03:39:54.867Z", "content": "Thanks for the clarification!
", "rawMarkdown": "Thanks for the clarification!" }, { "id": 1662204, "postDate": "2022-01-24T05:50:06.840Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 1654155, "postDate": "2022-01-18T08:02:04.647Z", "content": "Did public lb make the appropriate updates? If not, does it mean that when I add supplemental_train.csv to the training set, the public lb is no longer relevant for my model？
", "rawMarkdown": "Did public lb make the appropriate updates? If not, does it mean that when I add **supplemental_train.csv** to the training set, the public lb is no longer relevant for my model？", "replies": [ { "id": 1654900, "postDate": "2022-01-18T23:26:35.073Z", "content": "I assume the same that LB is already part of the supplemental_train.csv (which starts from 2021-09-21 00:01:00)… supplemental_train.csv was supposed to fill the gap between train.csv and test.csv for debugging purpose; however, it is not clear that supplemental_train.csv will always have no time gaps with the test set in the forecasting phase updates and there seem to be some outliers in the supplemental_train.csv as mentioned by @tomforbes, I would suggest to keep it out of the training set for now.
", "rawMarkdown": "I assume the same that LB is already part of the supplemental_train.csv (which starts from 2021-09-21 00:01:00)... supplemental_train.csv was supposed to fill the gap between train.csv and test.csv for debugging purpose; however, it is not clear that supplemental_train.csv will always have no time gaps with the test set in the forecasting phase updates and there seem to be some outliers in the supplemental_train.csv as mentioned by @tomforbes, I would suggest to keep it out of the training set for now." }, { "id": 1655996, "postDate": "2022-01-19T03:09:56.590Z", "content": "Thanks 😄, I just joined this comp and I now understand what LB is
", "rawMarkdown": "Thanks 😄, I just joined this comp and I now understand what LB is" } ] }, { "id": 1657254, "postDate": "2022-01-20T04:03:21.553Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1653199, "author_name": "Jose Antonio Alatorre", "author_url": "", "post_date": "2022-01-17T10:36:42.580000", "content": "Hi @sohier as mentioned by @tomforbes the new data has rows that are clearly wrong prices. For example, the price of Tron changed in 1-minute bar 6888% several times on October 29. The important fact is that the training set seems to be already curated filtering out any exchange or flash crashes while the supplemental not. These types of returns have a huge impact on modeling as are completely random and have several orders of magnitude. Should we assume that the final data set will include these type of price action?
", "votes": 7, "replies": [ { "id": 1653892, "author_name": "gyugyugyu", "author_url": "", "post_date": "2022-01-18T00:46:40.827000", "content": "Compared to Binance data, I can confidently say that this data contains errors. Now, what I wonder is if this error will be fixed soon, and the final supplementary data set includes such errors.
", "votes": 3, "replies": [] }, { "id": 1657115, "author_name": "Sohier Dane", "author_url": "", "post_date": "2022-01-19T23:14:01.040000", "content": "I've just uploaded a correction to the new data; it should be available to your notebooks soon.
", "votes": 5, "replies": [] }, { "id": 1657129, "author_name": "Yue Sun", "author_url": "", "post_date": "2022-01-19T23:46:10.507000", "content": "Thanks for the update!
", "votes": 1, "replies": [] } ] }, { "id": 1651166, "author_name": "Tom Forbes", "author_url": "", "post_date": "2022-01-15T15:39:03.813000", "content": "Hi @sohier, there appears to be some errors in the update which i have outlined here
", "votes": 8, "replies": [ { "id": 1652996, "author_name": "JonnydosSantos", "author_url": "", "post_date": "2022-01-17T06:51:42.187000", "content": "Many thanks for pointing this out, I experienced similar issues. Hopefully the Kaggle staff and organizers can give us some clarity on this.
", "votes": 0, "replies": [] }, { "id": 1654871, "author_name": "gmobaz", "author_url": "", "post_date": "2022-01-18T22:28:36.083000", "content": "Thank you very much, @tomforbes, @joseantonioalatorre and other kagglers! Three days ago you posted this valuable info. Unfortunately, nobody inside Kaggle seems to care..
", "votes": 0, "replies": [] }, { "id": 1654876, "author_name": "Sohier Dane", "author_url": "", "post_date": "2022-01-18T22:35:29.970000", "content": "@tomforbes thank you for flagging this, we are looking into the matter.
", "votes": 4, "replies": [] }, { "id": 1654884, "author_name": "gmobaz", "author_url": "", "post_date": "2022-01-18T22:57:43.320000", "content": "Thank you, @sohier and apologies for my previous claim. I see that with the added data, some targets can be updated also.
", "votes": 1, "replies": [] } ] }, { "id": 1651312, "author_name": "GUNER", "author_url": "", "post_date": "2022-01-15T17:11:11.970000", "content": "Why is this 'missing one minute' each month? Yes I am talking about the first minute of 4 months in the dataset?
", "votes": 1, "replies": [] }, { "id": 1650899, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2022-01-15T12:05:07.770000", "content": "Is there time interval between training data and test data ?
", "votes": 1, "replies": [ { "id": 1651322, "author_name": "GUNER", "author_url": "", "post_date": "2022-01-15T17:14:09.927000", "content": "In the evaluation phase, there will not be any data gap. So you can safely assume that your model will start predicting just after the last minute of data used in training… except the one minute that that is mentioned below.
", "votes": 2, "replies": [] }, { "id": 1652776, "author_name": "C4rl05/V", "author_url": "", "post_date": "2022-01-17T01:30:45.013000", "content": "Hello @abdulkadirguner, based on the original train and test datasets, I believe we are predicting for the next 3 months every minute into the future. Do you think this will be the case after the competition is close, trying to understand how much are we predicting / forecasting into the future?
", "votes": 1, "replies": [] }, { "id": 1652852, "author_name": "GUNER", "author_url": "", "post_date": "2022-01-17T03:55:55.723000", "content": "@cv12j0, you can say that. Competition ends beginning of February. Submitted notebooks will run against real data in the next 3 months period (until beginning of May). With one nuisance: this evaluation phase is divided into 2-week periods. So every two weeks, your notebooks will be assessed with data from the beginning of the test phase to the end of that two weeks. So if you are end of week 4, you will be assessed for the whole 4 weeks, and at the end of week 6, for the whole 6 weeks. Buttt, there is only one final score and that is your score for the whole 3-months of data, from start to finish at then end of week 12(end of month3).
", "votes": 1, "replies": [] }, { "id": 1652857, "author_name": "C4rl05/V", "author_url": "", "post_date": "2022-01-17T04:08:02.497000", "content": "Thanks, @abdulkadirguner trying to build lag features that can handle any future condition.
", "votes": 0, "replies": [] }, { "id": 1652864, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2022-01-17T04:18:09.900000", "content": "It seems that the current supplemental_train.csv file is completely different from train.csv. 👀
", "votes": 1, "replies": [] }, { "id": 1653990, "author_name": "C4rl05/V", "author_url": "", "post_date": "2022-01-18T03:13:05.367000", "content": "Yes @yuanzhezhou it is, but the main reason is to provide extra data, as they explain in the data description section
", "votes": 0, "replies": [] } ] }, { "id": 1654893, "author_name": "khyeh", "author_url": "", "post_date": "2022-01-18T23:18:12.417000", "content": "@sohier In the forecasting phase, since the supplemental_train.csv is already fixed, there will be time gaps between supplemental_train.csv and test after 1st update, am I understanding correctly?
\nI would also like to know for sure, but I believe the test will be rerun from the same starting point each update like:
\nweek 0-2
\nweek 0-4
\nweek 0-6
\n…..
\nweek 0-12 (final update)
This is a good point actually since either we time the notebook so that it is sufficient just to run for 2 weeks or the full 12 weeks. Feature engineering time will be limited for the latter case.
", "votes": 0, "replies": [] }, { "id": 1656213, "author_name": "khyeh", "author_url": "", "post_date": "2022-01-19T07:14:43.720000", "content": "week 0-12 will take a lot of time to compute then..
", "votes": 0, "replies": [] }, { "id": 1656807, "author_name": "Carlos Stein N Brito", "author_url": "", "post_date": "2022-01-19T16:32:01.233000", "content": "Each re-run will start from the same time point, with the final re-run (the only one that matters) including 3 months of data. The Public Leaderboard has similar amount of data and should be used to test your running time.
", "votes": 8, "replies": [] }, { "id": 1657237, "author_name": "khyeh", "author_url": "", "post_date": "2022-01-20T03:39:54.867000", "content": "Thanks for the clarification!
", "votes": 0, "replies": [] }, { "id": 1662204, "author_name": "", "author_url": "", "post_date": "2022-01-24T05:50:06.840000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1654155, "author_name": "AndyX", "author_url": "", "post_date": "2022-01-18T08:02:04.647000", "content": "Did public lb make the appropriate updates? If not, does it mean that when I add supplemental_train.csv to the training set, the public lb is no longer relevant for my model？
", "votes": 0, "replies": [ { "id": 1654900, "author_name": "khyeh", "author_url": "", "post_date": "2022-01-18T23:26:35.073000", "content": "I assume the same that LB is already part of the supplemental_train.csv (which starts from 2021-09-21 00:01:00)… supplemental_train.csv was supposed to fill the gap between train.csv and test.csv for debugging purpose; however, it is not clear that supplemental_train.csv will always have no time gaps with the test set in the forecasting phase updates and there seem to be some outliers in the supplemental_train.csv as mentioned by @tomforbes, I would suggest to keep it out of the training set for now.
", "votes": 0, "replies": [] }, { "id": 1655996, "author_name": "AndyX", "author_url": "", "post_date": "2022-01-19T03:09:56.590000", "content": "Thanks 😄, I just joined this comp and I now understand what LB is
", "votes": 0, "replies": [] } ] }, { "id": 1657254, "author_name": "", "author_url": "", "post_date": "2022-01-20T04:03:21.553000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1650142": "I'm happy to announce that we are releasing an interim update to the **supplemental_train.csv** file. This will enable you to get a head start on updating your models to incorporate data gathered since the competition began. It should also reduce the risk of error from any workarounds to handle the old placeholder data and provides an opportunity for you to test your models on this updated data before submissions close on 1 Feb. None of the other files will change. The update should be live within the next hour.\n\nThere will still be one more update to the **supplemental_train.csv** on the original schedule (after the submission deadline), which will cover the remainder of the period up until the submission deadline. Barring any missing data, the final copy of **supplemental_train.csv** will end right before the forecasting phase test set data begins.\n\nEdit: @tomforbes has identified some implausible values in the updated supplement. We are looking into the matter and will provide an update when we have more information.\n\nEdit 2: There was indeed some erroneous data coming from the data provider, which have been corrected and updated in the supplemental data file. The corrections should be available to your notebooks within the next hour. ", "1653199": "Hi @sohier as mentioned by @tomforbes the new data has rows that are clearly wrong prices. For example, the price of Tron changed in 1-minute bar 6888% several times on October 29. The important fact is that the training set seems to be already curated filtering out any exchange or flash crashes while the supplemental not. These types of returns have a huge impact on modeling as are completely random and have several orders of magnitude. Should we assume that the final data set will include these type of price action? ", "1651166": "Hi @sohier, there appears to be some errors in the update [which i have outlined here](https://www.kaggle.com/tomforbes/errors-in-supplemental-train-csv)", "1651312": "Why is this 'missing one minute' each month? Yes I am talking about the first minute of 4 months in the dataset?\n", "1650899": "Is there time interval between training data and test data ?", "1654893": "@sohier In the forecasting phase, since the supplemental_train.csv is already fixed, there will be time gaps between supplemental_train.csv and test after 1st update, am I understanding correctly?\n\n- forecast week 0: supplemental_train.csv up to week 0 (the final copy)\n- (1st update) forecast week 2: supplemental_train.csv up to week 0 + test.csv from week 0 to week 2\n- (2nd update) forecast week 4: supplemental_train.csv up to week 0 + test.csv from week 2 to week 4", "1654155": "Did public lb make the appropriate updates? If not, does it mean that when I add **supplemental_train.csv** to the training set, the public lb is no longer relevant for my model？", "1657254": "" } }