{
  "id": 343919,
  "title": "Using correlated input features for multi output regression",
  "url": "/competitions/g-research-crypto-forecasting/discussion/343919",
  "author_name": "",
  "post_date": "2022-08-13T02:06:56.155508800Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello,<br>\nI am fairly new to time series problems and especially to ones involving multi output regression. I am using this dataset to get started with these types of problems. <br>\nI tried to go through the tutorial notebook provided in the Competition's official documentation and also one notebook provided by the Gold Medalist. In both of these notebook they used upper shadow and lower shadow as features. The competition tutorial notebook also used log returns as well.<br>\nAs someone who is familiar with linear regression problems quite a bit, I know the key to obtain stable regression model parameters is to use uncorrelated or independent input features. But in current case looking at the nature of the input features they all look quite dependent or atleast derived from similar underlying data. Will not this be a problem?</p>\n<p>Thanks in advance for any help!</p>",
  "messages": [
    {
      "id": "1896626",
      "postDate": "08/13/2022 02:06:56",
      "content": "<p>Hello,<br>\nI am fairly new to time series problems and especially to ones involving multi output regression. I am using this dataset to get started with these types of problems. <br>\nI tried to go through the tutorial notebook provided in the Competition's official documentation and also one notebook provided by the Gold Medalist. In both of these notebook they used upper shadow and lower shadow as features. The competition tutorial notebook also used log returns as well.<br>\nAs someone who is familiar with linear regression problems quite a bit, I know the key to obtain stable regression model parameters is to use uncorrelated or independent input features. But in current case looking at the nature of the input features they all look quite dependent or atleast derived from similar underlying data. Will not this be a problem?</p>\n<p>Thanks in advance for any help!</p>",
      "rawMarkdown": "Hello,\nI am fairly new to time series problems and especially to ones involving multi output regression. I am using this dataset to get started with these types of problems. \nI tried to go through the tutorial notebook provided in the Competition's official documentation and also one notebook provided by the Gold Medalist. In both of these notebook they used upper shadow and lower shadow as features. The competition tutorial notebook also used log returns as well.\nAs someone who is familiar with linear regression problems quite a bit, I know the key to obtain stable regression model parameters is to use uncorrelated or independent input features. But in current case looking at the nature of the input features they all look quite dependent or atleast derived from similar underlying data. Will not this be a problem?\n\nThanks in advance for any help!",
      "votes": null
    },
    {
      "id": "1896859",
      "postDate": "08/13/2022 08:17:21",
      "content": "<p>Perfect independence between features is an assumption/ideal state, far apart from real life problems/ competition data. I suggest you may consider a metric like collinearity diagnostics and influence plots to evaluate your model. You may also consider transforming the features to perhaps reduce the correlation/ mutual information and use them better. <br>\nI hope this helps and wish you the best!</p>",
      "rawMarkdown": "Perfect independence between features is an assumption/ideal state, far apart from real life problems/ competition data. I suggest you may consider a metric like collinearity diagnostics and influence plots to evaluate your model. You may also consider transforming the features to perhaps reduce the correlation/ mutual information and use them better. \nI hope this helps and wish you the best!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1896859,
      "author_name": "ravi20076",
      "author_url": "",
      "post_date": "08/13/2022 08:17:21",
      "content": "<p>Perfect independence between features is an assumption/ideal state, far apart from real life problems/ competition data. I suggest you may consider a metric like collinearity diagnostics and influence plots to evaluate your model. You may also consider transforming the features to perhaps reduce the correlation/ mutual information and use them better. <br>\nI hope this helps and wish you the best!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1896626": "Hello,\nI am fairly new to time series problems and especially to ones involving multi output regression. I am using this dataset to get started with these types of problems. \nI tried to go through the tutorial notebook provided in the Competition's official documentation and also one notebook provided by the Gold Medalist. In both of these notebook they used upper shadow and lower shadow as features. The competition tutorial notebook also used log returns as well.\nAs someone who is familiar with linear regression problems quite a bit, I know the key to obtain stable regression model parameters is to use uncorrelated or independent input features. But in current case looking at the nature of the input features they all look quite dependent or atleast derived from similar underlying data. Will not this be a problem?\n\nThanks in advance for any help!",
    "1896859": "Perfect independence between features is an assumption/ideal state, far apart from real life problems/ competition data. I suggest you may consider a metric like collinearity diagnostics and influence plots to evaluate your model. You may also consider transforming the features to perhaps reduce the correlation/ mutual information and use them better. \nI hope this helps and wish you the best!"
  },
  "source": "meta"
}