{
  "id": 582473,
  "title": "What is your highest LB Score without DATA LEAK？(KFold is definitely a data leakage)",
  "url": "/competitions/drw-crypto-market-prediction/discussion/582473",
  "author_name": "Mr RRR",
  "post_date": "2025-05-31T14:21:06.497000",
  "votes": 0,
  "comment_count": 6,
  "views": 0,
  "content": "<p>The open source model with 0.1+ LB Score use KFold, which is a data leak. (Correct me if I am wrong, please)</p>\n<p>My currently highest LB Score without DATA LEAK is 0.076 using XGBoost model, what is yours?</p>",
  "messages": [
    {
      "id": 3215856,
      "postDate": "2025-06-02T18:15:56.240Z",
      "content": "<p><a href=\"https://www.kaggle.com/larrylin666\" target=\"_blank\">@larrylin666</a> How did you get to 0.088? Did you end up doing KFold? </p>",
      "rawMarkdown": "@larrylin666 How did you get to 0.088? Did you end up doing KFold? ",
      "votes": 1,
      "replies": [
        {
          "id": 3216326,
          "postDate": "2025-06-03T12:23:50.500Z",
          "content": "<p>I would say no data leak</p>",
          "rawMarkdown": "I would say no data leak"
        }
      ]
    },
    {
      "id": 3214642,
      "postDate": "2025-05-31T23:38:34.010Z",
      "content": "<p>KFold is a leak on the validation data but not really the leaderboard. The reason is because you are training on future datapoints/data that are much closer in time to the validation data than what they would be on the LB data so you get a validation score that is not representative of how your model actually generalizes</p>",
      "rawMarkdown": "KFold is a leak on the validation data but not really the leaderboard. The reason is because you are training on future datapoints/data that are much closer in time to the validation data than what they would be on the LB data so you get a validation score that is not representative of how your model actually generalizes",
      "votes": 1,
      "replies": [
        {
          "id": 3214656,
          "postDate": "2025-06-01T00:14:17.883Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 3214669,
          "postDate": "2025-06-01T00:30:57.383Z",
          "content": "<p>I see, but as a timeseries problem, it is not scientific to solve it this this…</p>",
          "rawMarkdown": "I see, but as a timeseries problem, it is not scientific to solve it this this...",
          "replies": [
            {
              "id": 3214675,
              "postDate": "2025-06-01T00:53:11.937Z",
              "content": "<p>Yes I agree kfold isn’t a good way to validate models in this challenge. I was just saying that it’s not a leak of lb</p>",
              "rawMarkdown": "Yes I agree kfold isn’t a good way to validate models in this challenge. I was just saying that it’s not a leak of lb"
            }
          ]
        }
      ]
    },
    {
      "id": 3214431,
      "postDate": "2025-05-31T14:21:06.497Z",
      "content": "<p>The open source model with 0.1+ LB Score use KFold, which is a data leak. (Correct me if I am wrong, please)</p>\n<p>My currently highest LB Score without DATA LEAK is 0.076 using XGBoost model, what is yours?</p>",
      "rawMarkdown": "The open source model with 0.1+ LB Score use KFold, which is a data leak. (Correct me if I am wrong, please)\n\nMy currently highest LB Score without DATA LEAK is 0.076 using XGBoost model, what is yours?",
      "votes": -1
    }
  ],
  "comments": [
    {
      "id": 3215856,
      "author_name": "byunjins",
      "author_url": "",
      "post_date": "2025-06-02T18:15:56.240000",
      "content": "<p><a href=\"https://www.kaggle.com/larrylin666\" target=\"_blank\">@larrylin666</a> How did you get to 0.088? Did you end up doing KFold? </p>",
      "votes": 1,
      "replies": [
        {
          "id": 3216326,
          "author_name": "Mr RRR",
          "author_url": "",
          "post_date": "2025-06-03T12:23:50.500000",
          "content": "<p>I would say no data leak</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3214642,
      "author_name": "snehal",
      "author_url": "",
      "post_date": "2025-05-31T23:38:34.010000",
      "content": "<p>KFold is a leak on the validation data but not really the leaderboard. The reason is because you are training on future datapoints/data that are much closer in time to the validation data than what they would be on the LB data so you get a validation score that is not representative of how your model actually generalizes</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3214656,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-01T00:14:17.883000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3214669,
          "author_name": "Mr RRR",
          "author_url": "",
          "post_date": "2025-06-01T00:30:57.383000",
          "content": "<p>I see, but as a timeseries problem, it is not scientific to solve it this this…</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3214675,
              "author_name": "snehal",
              "author_url": "",
              "post_date": "2025-06-01T00:53:11.937000",
              "content": "<p>Yes I agree kfold isn’t a good way to validate models in this challenge. I was just saying that it’s not a leak of lb</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3215856": "@larrylin666 How did you get to 0.088? Did you end up doing KFold? ",
    "3214642": "KFold is a leak on the validation data but not really the leaderboard. The reason is because you are training on future datapoints/data that are much closer in time to the validation data than what they would be on the LB data so you get a validation score that is not representative of how your model actually generalizes",
    "3214431": "The open source model with 0.1+ LB Score use KFold, which is a data leak. (Correct me if I am wrong, please)\n\nMy currently highest LB Score without DATA LEAK is 0.076 using XGBoost model, what is yours?"
  }
}