{
  "id": 545430,
  "title": "Bad dates?",
  "url": "/competitions/jane-street-real-time-market-data-forecasting/discussion/545430",
  "author_name": "",
  "post_date": "2024-11-10T05:16:41.933818600Z",
  "votes": 7,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I saw some of the public notebooks were dropping a large portion of the training set because of memory limitations. I also tried to filter out some dates before <code>time_id</code> count increase (I can't remember the exact point where it happens) and my validation score deviated within a range of +-0.0005. It seems like a small subset of dates can have huge influence on the score. I think it would be better to use all training data and bagging to build diverse trees.  </p>",
  "messages": [
    {
      "id": "3041266",
      "postDate": "11/10/2024 05:16:41",
      "content": "<p>I saw some of the public notebooks were dropping a large portion of the training set because of memory limitations. I also tried to filter out some dates before <code>time_id</code> count increase (I can't remember the exact point where it happens) and my validation score deviated within a range of +-0.0005. It seems like a small subset of dates can have huge influence on the score. I think it would be better to use all training data and bagging to build diverse trees.  </p>",
      "rawMarkdown": "I saw some of the public notebooks were dropping a large portion of the training set because of memory limitations. I also tried to filter out some dates before `time_id` count increase (I can't remember the exact point where it happens) and my validation score deviated within a range of +-0.0005. It seems like a small subset of dates can have huge influence on the score. I think it would be better to use all training data and bagging to build diverse trees.",
      "votes": null
    },
    {
      "id": "3041292",
      "postDate": "11/10/2024 06:10:03",
      "content": "<p>Time id changes on date 698. I agree that the choice of dates can have a profound impact on the CV score and even on the leaderboard. A lot of the participants are not able to get any form of CV-LB correlation as well <a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a> </p>",
      "rawMarkdown": "Time id changes on date 698. I agree that the choice of dates can have a profound impact on the CV score and even on the leaderboard. A lot of the participants are not able to get any form of CV-LB correlation as well @gunesevitan",
      "votes": null
    },
    {
      "id": "3041295",
      "postDate": "11/10/2024 06:16:49",
      "content": "<p>Thanks, yeah it was date 698. That is similar to my experience so far.</p>",
      "rawMarkdown": "Thanks, yeah it was date 698. That is similar to my experience so far.",
      "votes": null
    },
    {
      "id": "3042961",
      "postDate": "11/12/2024 00:57:07",
      "content": "<p>Hi, </p>\n<p>What is considered a 'bad' date, and why date id 698?</p>",
      "rawMarkdown": "Hi, \n\nWhat is considered a 'bad' date, and why date id 698?",
      "votes": null
    },
    {
      "id": "3043194",
      "postDate": "11/12/2024 06:44:14",
      "content": "<p>After date id 698, number of time ids changes to a new state. I tried dropping some dates before that and my score both got better and worse depending on the dates I was dropping. I was thinking some of the dates might have bad influence but I haven't dig deep into that yet.</p>",
      "rawMarkdown": "After date id 698, number of time ids changes to a new state. I tried dropping some dates before that and my score both got better and worse depending on the dates I was dropping. I was thinking some of the dates might have bad influence but I haven't dig deep into that yet.",
      "votes": null
    },
    {
      "id": "3043408",
      "postDate": "11/12/2024 12:03:11",
      "content": "<p>Isn't it date_id 676-677 where the breaking point is? <br>\nWith date_id 676 included we have 849 time_ids and with 677 onward there are 968 time_ids </p>",
      "rawMarkdown": "Isn't it date_id 676-677 where the breaking point is? \nWith date_id 676 included we have 849 time_ids and with 677 onward there are 968 time_ids",
      "votes": null
    },
    {
      "id": "3043422",
      "postDate": "11/12/2024 12:14:08",
      "content": "<p>exactly. <br>\nThats why i was confused with the number 698 given above</p>",
      "rawMarkdown": "exactly. \nThats why i was confused with the number 698 given above",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3041292,
      "author_name": "ravi20076",
      "author_url": "",
      "post_date": "11/10/2024 06:10:03",
      "content": "<p>Time id changes on date 698. I agree that the choice of dates can have a profound impact on the CV score and even on the leaderboard. A lot of the participants are not able to get any form of CV-LB correlation as well <a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a> </p>",
      "votes": null,
      "replies": [
        {
          "id": 3041295,
          "author_name": "gunesevitan",
          "author_url": "",
          "post_date": "11/10/2024 06:16:49",
          "content": "<p>Thanks, yeah it was date 698. That is similar to my experience so far.</p>",
          "votes": null,
          "replies": [
            {
              "id": 3042961,
              "author_name": "hashirsherwani",
              "author_url": "",
              "post_date": "11/12/2024 00:57:07",
              "content": "<p>Hi, </p>\n<p>What is considered a 'bad' date, and why date id 698?</p>",
              "votes": null,
              "replies": [
                {
                  "id": 3043194,
                  "author_name": "gunesevitan",
                  "author_url": "",
                  "post_date": "11/12/2024 06:44:14",
                  "content": "<p>After date id 698, number of time ids changes to a new state. I tried dropping some dates before that and my score both got better and worse depending on the dates I was dropping. I was thinking some of the dates might have bad influence but I haven't dig deep into that yet.</p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        },
        {
          "id": 3043408,
          "author_name": "eu1234",
          "author_url": "",
          "post_date": "11/12/2024 12:03:11",
          "content": "<p>Isn't it date_id 676-677 where the breaking point is? <br>\nWith date_id 676 included we have 849 time_ids and with 677 onward there are 968 time_ids </p>",
          "votes": null,
          "replies": [
            {
              "id": 3043422,
              "author_name": "hashirsherwani",
              "author_url": "",
              "post_date": "11/12/2024 12:14:08",
              "content": "<p>exactly. <br>\nThats why i was confused with the number 698 given above</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3041266": "I saw some of the public notebooks were dropping a large portion of the training set because of memory limitations. I also tried to filter out some dates before `time_id` count increase (I can't remember the exact point where it happens) and my validation score deviated within a range of +-0.0005. It seems like a small subset of dates can have huge influence on the score. I think it would be better to use all training data and bagging to build diverse trees.",
    "3041292": "Time id changes on date 698. I agree that the choice of dates can have a profound impact on the CV score and even on the leaderboard. A lot of the participants are not able to get any form of CV-LB correlation as well @gunesevitan",
    "3041295": "Thanks, yeah it was date 698. That is similar to my experience so far.",
    "3042961": "Hi, \n\nWhat is considered a 'bad' date, and why date id 698?",
    "3043194": "After date id 698, number of time ids changes to a new state. I tried dropping some dates before that and my score both got better and worse depending on the dates I was dropping. I was thinking some of the dates might have bad influence but I haven't dig deep into that yet.",
    "3043408": "Isn't it date_id 676-677 where the breaking point is? \nWith date_id 676 included we have 849 time_ids and with 677 onward there are 968 time_ids",
    "3043422": "exactly. \nThats why i was confused with the number 698 given above"
  },
  "source": "meta"
}