{
  "id": 586789,
  "title": "Time based features",
  "url": "/competitions/drw-crypto-market-prediction/discussion/586789",
  "author_name": "",
  "post_date": "2025-06-28T10:55:16.035860Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I was trying to understand if past values ​​can give an indication of future values ​​to see for example if the shifted values ​​of the features can be used as feature engineering trying for example to evaluate autocorrelation of target label in train dataset; this approach to work however implies that in the test dataset all the rows are temporally consecutive but it is not so because the timestamps in the test dataset are replaced by IDs and above all the rows in the test set are shuffled so basically it is not possible to create new features based on time since in the test dataset the data are not ordered by time (timestamp: To prevent future peeking, all timestamps are masked, shuffled, and replaced with a unique ID.). What do you think?</p>",
  "messages": [
    {
      "id": "3234808",
      "postDate": "06/28/2025 10:55:16",
      "content": "<p>I was trying to understand if past values ​​can give an indication of future values ​​to see for example if the shifted values ​​of the features can be used as feature engineering trying for example to evaluate autocorrelation of target label in train dataset; this approach to work however implies that in the test dataset all the rows are temporally consecutive but it is not so because the timestamps in the test dataset are replaced by IDs and above all the rows in the test set are shuffled so basically it is not possible to create new features based on time since in the test dataset the data are not ordered by time (timestamp: To prevent future peeking, all timestamps are masked, shuffled, and replaced with a unique ID.). What do you think?</p>",
      "rawMarkdown": "I was trying to understand if past values ​​can give an indication of future values ​​to see for example if the shifted values ​​of the features can be used as feature engineering trying for example to evaluate autocorrelation of target label in train dataset; this approach to work however implies that in the test dataset all the rows are temporally consecutive but it is not so because the timestamps in the test dataset are replaced by IDs and above all the rows in the test set are shuffled so basically it is not possible to create new features based on time since in the test dataset the data are not ordered by time (timestamp: To prevent future peeking, all timestamps are masked, shuffled, and replaced with a unique ID.). What do you think?",
      "votes": null
    },
    {
      "id": "3235364",
      "postDate": "06/29/2025 05:56:12",
      "content": "<p>I also experimented with shifted values ​​of features, it did not improve the score!</p>",
      "rawMarkdown": "I also experimented with shifted values ​​of features, it did not improve the score!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3235364,
      "author_name": "itasps",
      "author_url": "",
      "post_date": "06/29/2025 05:56:12",
      "content": "<p>I also experimented with shifted values ​​of features, it did not improve the score!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3234808": "I was trying to understand if past values ​​can give an indication of future values ​​to see for example if the shifted values ​​of the features can be used as feature engineering trying for example to evaluate autocorrelation of target label in train dataset; this approach to work however implies that in the test dataset all the rows are temporally consecutive but it is not so because the timestamps in the test dataset are replaced by IDs and above all the rows in the test set are shuffled so basically it is not possible to create new features based on time since in the test dataset the data are not ordered by time (timestamp: To prevent future peeking, all timestamps are masked, shuffled, and replaced with a unique ID.). What do you think?",
    "3235364": "I also experimented with shifted values ​​of features, it did not improve the score!"
  },
  "source": "meta"
}