{
  "id": 66205,
  "title": "RMSLE: Predicting higher values is better (than lower)?",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/66205",
  "author_name": "",
  "post_date": "2018-09-19T08:05:33.936519500Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I did a quick analysis on RMSLE in <a href=\"https://www.kaggle.com/krisho007/rmsle-analysis\">this kerne</a>l. My conclusion is that you are better off predicting more than the actual revenue, rather than less than actual revenue. Feel free to share your thoughts.  </p>",
  "messages": [
    {
      "id": "389778",
      "postDate": "09/19/2018 08:05:33",
      "content": "<p>I did a quick analysis on RMSLE in <a href=\"https://www.kaggle.com/krisho007/rmsle-analysis\">this kerne</a>l. My conclusion is that you are better off predicting more than the actual revenue, rather than less than actual revenue. Feel free to share your thoughts.  </p>",
      "rawMarkdown": "I did a quick analysis on RMSLE in [this kerne][1]l. My conclusion is that you are better off predicting more than the actual revenue, rather than less than actual revenue. Feel free to share your thoughts.  \n\n\n  [1]: https://www.kaggle.com/krisho007/rmsle-analysis",
      "votes": null
    },
    {
      "id": "389833",
      "postDate": "09/19/2018 09:40:37",
      "content": "<p>As mentioned in other threads in the discussions, this could be caused by either 1) existence of some big target values just in the public lb or 2) the difference between train and test data (set of target values in test consists of considerably greater values).\nIf the first case is true, using sum of log means overfitting (aka doom)! otherwise it is like some sort of post processing to lower the error (i think many top teams are trying different post processing methods, it's like gambling ;) ).  Maybe using other metrics that are less vulnerable to outliers would have been better here.</p>",
      "rawMarkdown": "As mentioned in other threads in the discussions, this could be caused by either 1) existence of some big target values just in the public lb or 2) the difference between train and test data (set of target values in test consists of considerably greater values).\nIf the first case is true, using sum of log means overfitting (aka doom)! otherwise it is like some sort of post processing to lower the error (i think many top teams are trying different post processing methods, it's like gambling ;) ).  Maybe using other metrics that are less vulnerable to outliers would have been better here.",
      "votes": null
    },
    {
      "id": "392849",
      "postDate": "09/24/2018 13:28:21",
      "content": "<p>Yes 'cause it's better not to miss a big fish and be a bit overly positive ) While showing special promotion to a few of mere mortals won't hurt same much as losing a valuable customer.</p>",
      "rawMarkdown": "Yes 'cause it's better not to miss a big fish and be a bit overly positive ) While showing special promotion to a few of mere mortals won't hurt same much as losing a valuable customer.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 389833,
      "author_name": "kainsama",
      "author_url": "",
      "post_date": "09/19/2018 09:40:37",
      "content": "<p>As mentioned in other threads in the discussions, this could be caused by either 1) existence of some big target values just in the public lb or 2) the difference between train and test data (set of target values in test consists of considerably greater values).\nIf the first case is true, using sum of log means overfitting (aka doom)! otherwise it is like some sort of post processing to lower the error (i think many top teams are trying different post processing methods, it's like gambling ;) ).  Maybe using other metrics that are less vulnerable to outliers would have been better here.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 392849,
      "author_name": "fingoldo",
      "author_url": "",
      "post_date": "09/24/2018 13:28:21",
      "content": "<p>Yes 'cause it's better not to miss a big fish and be a bit overly positive ) While showing special promotion to a few of mere mortals won't hurt same much as losing a valuable customer.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "389778": "I did a quick analysis on RMSLE in [this kerne][1]l. My conclusion is that you are better off predicting more than the actual revenue, rather than less than actual revenue. Feel free to share your thoughts.  \n\n\n  [1]: https://www.kaggle.com/krisho007/rmsle-analysis",
    "389833": "As mentioned in other threads in the discussions, this could be caused by either 1) existence of some big target values just in the public lb or 2) the difference between train and test data (set of target values in test consists of considerably greater values).\nIf the first case is true, using sum of log means overfitting (aka doom)! otherwise it is like some sort of post processing to lower the error (i think many top teams are trying different post processing methods, it's like gambling ;) ).  Maybe using other metrics that are less vulnerable to outliers would have been better here.",
    "392849": "Yes 'cause it's better not to miss a big fish and be a bit overly positive ) While showing special promotion to a few of mere mortals won't hurt same much as losing a valuable customer."
  },
  "source": "meta"
}