{
  "id": 5432,
  "title": "training data for station CLAY(34.7,-95.3) from May-July,1999",
  "url": "/competitions/ams-2014-solar-energy-prediction-contest/discussion/5432",
  "author_name": "",
  "post_date": "2013-08-17T23:36:20.020Z",
  "votes": null,
  "comment_count": 2,
  "views": 1055,
  "content": "<p>Just some odds.&nbsp;The data shows a flat line with value&nbsp;12320768.</p>\n<p>Wonder what happened? Funny that it is a non zero value, if the site was down.</p>",
  "messages": [
    {
      "id": "28880",
      "postDate": "08/17/2013 23:36:20",
      "content": "<p>Just some odds.&nbsp;The data shows a flat line with value&nbsp;12320768.</p>\n<p>Wonder what happened? Funny that it is a non zero value, if the site was down.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28885",
      "postDate": "08/18/2013 01:14:27",
      "content": "<p>CHIC from early 1994 also flatlined at 1336944.&nbsp; Same with OKMU at 4980736 near the end and some others. You can find quite a few of these actually if you search for numbers that are not divisible by 100. They're all unique by station however. The numbers that were not divisible by 100 were always peculiar to me and I was meaning to ask the admin, but never had time to investigate fully, but those are my brief mental notes.</p>\n<p>&nbsp;</p>\n<p>I'd like it if the admin can address this issue please.</p>\n<p>&nbsp;</p>\n<p>EDIT: Quick check shows there are 3334 of these numbers that are not divisible by 100. With only 2 exceptions, if a number that is not divisible by 100 appears again in the station's training data, it is the same number even if it isn't continuously flat-lining. Meaning, except for 2 exceptions I see, numbers not divisible by 100 are unique to station.</p>\n<p>I could be wrong by searching based on numbers not divisible by 100, but there's quite a few similarities between those numbers and the flat-lining described by Peter.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28897",
      "postDate": "08/18/2013 15:12:00",
      "content": "<p>The Mesonet does experience station outages resulting in missing data. Because Kaggle's evaluation system does not handle missing values, I applied two data filling schemes to eliminate the gaps. For some cases I did a forward fill of the data, and for others I used a distance weighted interpolation scheme. The same procedures were applied to the training and testing data. The gaps should all be small enough and the fills close enough to the spirit of the true data that it should not have a significant impact on the results. If you are particularly concerned, try only training on days without the fill in values and see how much it affects your submission results.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 28885,
      "author_name": "duffman",
      "author_url": "",
      "post_date": "08/18/2013 01:14:27",
      "content": "<p>CHIC from early 1994 also flatlined at 1336944.&nbsp; Same with OKMU at 4980736 near the end and some others. You can find quite a few of these actually if you search for numbers that are not divisible by 100. They're all unique by station however. The numbers that were not divisible by 100 were always peculiar to me and I was meaning to ask the admin, but never had time to investigate fully, but those are my brief mental notes.</p>\n<p>&nbsp;</p>\n<p>I'd like it if the admin can address this issue please.</p>\n<p>&nbsp;</p>\n<p>EDIT: Quick check shows there are 3334 of these numbers that are not divisible by 100. With only 2 exceptions, if a number that is not divisible by 100 appears again in the station's training data, it is the same number even if it isn't continuously flat-lining. Meaning, except for 2 exceptions I see, numbers not divisible by 100 are unique to station.</p>\n<p>I could be wrong by searching based on numbers not divisible by 100, but there's quite a few similarities between those numbers and the flat-lining described by Peter.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 28897,
      "author_name": "djgagne",
      "author_url": "",
      "post_date": "08/18/2013 15:12:00",
      "content": "<p>The Mesonet does experience station outages resulting in missing data. Because Kaggle's evaluation system does not handle missing values, I applied two data filling schemes to eliminate the gaps. For some cases I did a forward fill of the data, and for others I used a distance weighted interpolation scheme. The same procedures were applied to the training and testing data. The gaps should all be small enough and the fills close enough to the spirit of the true data that it should not have a significant impact on the results. If you are particularly concerned, try only training on days without the fill in values and see how much it affects your submission results.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "28880": "",
    "28885": "",
    "28897": ""
  },
  "source": "meta"
}