{
  "id": 303821,
  "title": "Do we train on full train.csv or only a part ie: data before 2021-06-13 00:00:00 for submission ",
  "url": "/competitions/g-research-crypto-forecasting/discussion/303821",
  "author_name": "",
  "post_date": "2022-01-29T14:18:06.186548100Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello all , </p>\n<p>For final submission do we train on the whole train dataset ie upto 2021-09 or just train on data before 2021-06-13 ? Also how to cross validate , do i use a simple time series split of sklearn or others ?<br>\nThanks</p>",
  "messages": [
    {
      "id": "1668318",
      "postDate": "01/29/2022 14:18:06",
      "content": "<p>Hello all , </p>\n<p>For final submission do we train on the whole train dataset ie upto 2021-09 or just train on data before 2021-06-13 ? Also how to cross validate , do i use a simple time series split of sklearn or others ?<br>\nThanks</p>",
      "rawMarkdown": "Hello all , \n\nFor final submission do we train on the whole train dataset ie upto 2021-09 or just train on data before 2021-06-13 ? Also how to cross validate , do i use a simple time series split of sklearn or others ?\nThanks",
      "votes": null
    },
    {
      "id": "1668571",
      "postDate": "01/29/2022 18:26:29",
      "content": "<p>Don't use full data for \"leaderboard submission\", but you definitely should for your final submission. <br>\nYou should only use training data (including holdout validation for early stopping) that come before the test. For validation, pretend the test set is at a certain time and use all the data before it to create a model. I've never used sklearn time series split but I'm sure that is how they do it. </p>",
      "rawMarkdown": "Don't use full data for \"leaderboard submission\", but you definitely should for your final submission. \nYou should only use training data (including holdout validation for early stopping) that come before the test. For validation, pretend the test set is at a certain time and use all the data before it to create a model. I've never used sklearn time series split but I'm sure that is how they do it.",
      "votes": null
    },
    {
      "id": "1668700",
      "postDate": "01/29/2022 23:44:56",
      "content": "<p>Hello, <a href=\"https://www.kaggle.com/ptrikp\" target=\"_blank\">@ptrikp</a>. There are only a few days left; make sure you incorporate all the new data released by the competition host for your final model.</p>",
      "rawMarkdown": "Hello, @ptrikp. There are only a few days left; make sure you incorporate all the new data released by the competition host for your final model.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1668571,
      "author_name": "chrisrichardmiles",
      "author_url": "",
      "post_date": "01/29/2022 18:26:29",
      "content": "<p>Don't use full data for \"leaderboard submission\", but you definitely should for your final submission. <br>\nYou should only use training data (including holdout validation for early stopping) that come before the test. For validation, pretend the test set is at a certain time and use all the data before it to create a model. I've never used sklearn time series split but I'm sure that is how they do it. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1668700,
      "author_name": "cv13j0",
      "author_url": "",
      "post_date": "01/29/2022 23:44:56",
      "content": "<p>Hello, <a href=\"https://www.kaggle.com/ptrikp\" target=\"_blank\">@ptrikp</a>. There are only a few days left; make sure you incorporate all the new data released by the competition host for your final model.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1668318": "Hello all , \n\nFor final submission do we train on the whole train dataset ie upto 2021-09 or just train on data before 2021-06-13 ? Also how to cross validate , do i use a simple time series split of sklearn or others ?\nThanks",
    "1668571": "Don't use full data for \"leaderboard submission\", but you definitely should for your final submission. \nYou should only use training data (including holdout validation for early stopping) that come before the test. For validation, pretend the test set is at a certain time and use all the data before it to create a model. I've never used sklearn time series split but I'm sure that is how they do it.",
    "1668700": "Hello, @ptrikp. There are only a few days left; make sure you incorporate all the new data released by the competition host for your final model."
  },
  "source": "meta"
}