{
  "id": 492074,
  "title": "Should we do K-Fold Cross Validation?",
  "url": "/competitions/home-credit-credit-risk-model-stability/discussion/492074",
  "author_name": "",
  "post_date": "2024-04-08T14:36:35.478566100Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Even in my current code, I am performing Stratified Group KFold Cross Validation with Shuffle=False.<br>\nHowever, I wonder if I do this, I will be using future data to make a certain prediction.<br>\nOr is there no problem because even though it is time series data, it is aggregated by CaseID?<br>\nWould love to hear your guys thoughts.</p>",
  "messages": [
    {
      "id": "2741680",
      "postDate": "04/08/2024 14:36:35",
      "content": "<p>Even in my current code, I am performing Stratified Group KFold Cross Validation with Shuffle=False.<br>\nHowever, I wonder if I do this, I will be using future data to make a certain prediction.<br>\nOr is there no problem because even though it is time series data, it is aggregated by CaseID?<br>\nWould love to hear your guys thoughts.</p>",
      "rawMarkdown": "Even in my current code, I am performing Stratified Group KFold Cross Validation with Shuffle=False.\nHowever, I wonder if I do this, I will be using future data to make a certain prediction.\nOr is there no problem because even though it is time series data, it is aggregated by CaseID?\nWould love to hear your guys thoughts.",
      "votes": null
    },
    {
      "id": "2742043",
      "postDate": "04/08/2024 17:54:21",
      "content": "<p>u can test these by further dividing the train data into train and test data, for example we have 92 week of data, u can use, first 50 weeks to train and see how it performs on the remaining week. this will give u better picture, that how the current strategy is useful or not in predicting over future week data. in the training part of first  50 weeks u can try k fold or any other u want to experiment to train. The strategy you find more useful,  can use it over entire train data to make the final model.</p>",
      "rawMarkdown": "u can test these by further dividing the train data into train and test data, for example we have 92 week of data, u can use, first 50 weeks to train and see how it performs on the remaining week. this will give u better picture, that how the current strategy is useful or not in predicting over future week data. in the training part of first  50 weeks u can try k fold or any other u want to experiment to train. The strategy you find more useful,  can use it over entire train data to make the final model.",
      "votes": null
    },
    {
      "id": "2749196",
      "postDate": "04/12/2024 22:35:42",
      "content": "<p>I have not joined this comp nor looked at the data. However if the data is time series, you will want to build validation schemes based on time. In other words, we train on on period of time and predict a future period of time.</p>",
      "rawMarkdown": "I have not joined this comp nor looked at the data. However if the data is time series, you will want to build validation schemes based on time. In other words, we train on on period of time and predict a future period of time.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2742043,
      "author_name": "shreyas9181",
      "author_url": "",
      "post_date": "04/08/2024 17:54:21",
      "content": "<p>u can test these by further dividing the train data into train and test data, for example we have 92 week of data, u can use, first 50 weeks to train and see how it performs on the remaining week. this will give u better picture, that how the current strategy is useful or not in predicting over future week data. in the training part of first  50 weeks u can try k fold or any other u want to experiment to train. The strategy you find more useful,  can use it over entire train data to make the final model.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2749196,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "04/12/2024 22:35:42",
      "content": "<p>I have not joined this comp nor looked at the data. However if the data is time series, you will want to build validation schemes based on time. In other words, we train on on period of time and predict a future period of time.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2741680": "Even in my current code, I am performing Stratified Group KFold Cross Validation with Shuffle=False.\nHowever, I wonder if I do this, I will be using future data to make a certain prediction.\nOr is there no problem because even though it is time series data, it is aggregated by CaseID?\nWould love to hear your guys thoughts.",
    "2742043": "u can test these by further dividing the train data into train and test data, for example we have 92 week of data, u can use, first 50 weeks to train and see how it performs on the remaining week. this will give u better picture, that how the current strategy is useful or not in predicting over future week data. in the training part of first  50 weeks u can try k fold or any other u want to experiment to train. The strategy you find more useful,  can use it over entire train data to make the final model.",
    "2749196": "I have not joined this comp nor looked at the data. However if the data is time series, you will want to build validation schemes based on time. In other words, we train on on period of time and predict a future period of time."
  },
  "source": "meta"
}