{
  "id": 300880,
  "title": "train_supplement & possible production issues",
  "url": "/competitions/g-research-crypto-forecasting/discussion/300880",
  "author_name": "",
  "post_date": "2022-01-14T17:39:30.452099900Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I just started the competition and i'am scratching my head about the train_supplement data.  <br>\nI understand that at evaluation-time  train, train_supplement and test (from API) will be contiguous. That way i could use the train data for feature engineering and for fitting timeseries models \"offline\" and use the train_supplement at evaluation time to build features and to predict the fetched test-data.  Unfortunately the information about the train_supplement is still quite \"sparse\".  Can we expect the train_supplement to \"interconnect smoothly\" with the train and the test data ? Will the data quality be \"goodish\"?  <br>\nI ask, because I see a high potential for possible \"production issues\" here.  Or did i get the whole thing wrong ?      </p>",
  "messages": [
    {
      "id": "1649938",
      "postDate": "01/14/2022 17:39:30",
      "content": "<p>I just started the competition and i'am scratching my head about the train_supplement data.  <br>\nI understand that at evaluation-time  train, train_supplement and test (from API) will be contiguous. That way i could use the train data for feature engineering and for fitting timeseries models \"offline\" and use the train_supplement at evaluation time to build features and to predict the fetched test-data.  Unfortunately the information about the train_supplement is still quite \"sparse\".  Can we expect the train_supplement to \"interconnect smoothly\" with the train and the test data ? Will the data quality be \"goodish\"?  <br>\nI ask, because I see a high potential for possible \"production issues\" here.  Or did i get the whole thing wrong ?      </p>",
      "rawMarkdown": "I just started the competition and i'am scratching my head about the train_supplement data.  \nI understand that at evaluation-time  train, train_supplement and test (from API) will be contiguous. That way i could use the train data for feature engineering and for fitting timeseries models \"offline\" and use the train_supplement at evaluation time to build features and to predict the fetched test-data.  Unfortunately the information about the train_supplement is still quite \"sparse\".  Can we expect the train_supplement to \"interconnect smoothly\" with the train and the test data ? Will the data quality be \"goodish\"?  \nI ask, because I see a high potential for possible \"production issues\" here.  Or did i get the whole thing wrong ?",
      "votes": null
    },
    {
      "id": "1650042",
      "postDate": "01/14/2022 19:13:10",
      "content": "<p>Yes, this is an excellent question.</p>\n<p>At this point also the dates in the supplemental_train overlap with the train. As such, there is no way to test concatenation, etc.</p>",
      "rawMarkdown": "Yes, this is an excellent question.\n\nAt this point also the dates in the supplemental_train overlap with the train. As such, there is no way to test concatenation, etc.",
      "votes": null
    },
    {
      "id": "1651648",
      "postDate": "01/15/2022 23:59:04",
      "content": "<p>Just sah an update here: <a href=\"https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/300906\" target=\"_blank\">https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/300906</a> <br>\nBut Tom Forbes already found some issues. Basically i still consider the supplement data somehow a blackbox. The thing is we have to decide for either a \"not too good but super-robust\" or a \"really good with the risk of a total fail in the evaluation\" approach. </p>",
      "rawMarkdown": "Just sah an update here: https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/300906 \nBut Tom Forbes already found some issues. Basically i still consider the supplement data somehow a blackbox. The thing is we have to decide for either a \"not too good but super-robust\" or a \"really good with the risk of a total fail in the evaluation\" approach.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1650042,
      "author_name": "marketneutral",
      "author_url": "",
      "post_date": "01/14/2022 19:13:10",
      "content": "<p>Yes, this is an excellent question.</p>\n<p>At this point also the dates in the supplemental_train overlap with the train. As such, there is no way to test concatenation, etc.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1651648,
      "author_name": "wysiati",
      "author_url": "",
      "post_date": "01/15/2022 23:59:04",
      "content": "<p>Just sah an update here: <a href=\"https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/300906\" target=\"_blank\">https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/300906</a> <br>\nBut Tom Forbes already found some issues. Basically i still consider the supplement data somehow a blackbox. The thing is we have to decide for either a \"not too good but super-robust\" or a \"really good with the risk of a total fail in the evaluation\" approach. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1649938": "I just started the competition and i'am scratching my head about the train_supplement data.  \nI understand that at evaluation-time  train, train_supplement and test (from API) will be contiguous. That way i could use the train data for feature engineering and for fitting timeseries models \"offline\" and use the train_supplement at evaluation time to build features and to predict the fetched test-data.  Unfortunately the information about the train_supplement is still quite \"sparse\".  Can we expect the train_supplement to \"interconnect smoothly\" with the train and the test data ? Will the data quality be \"goodish\"?  \nI ask, because I see a high potential for possible \"production issues\" here.  Or did i get the whole thing wrong ?",
    "1650042": "Yes, this is an excellent question.\n\nAt this point also the dates in the supplemental_train overlap with the train. As such, there is no way to test concatenation, etc.",
    "1651648": "Just sah an update here: https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/300906 \nBut Tom Forbes already found some issues. Basically i still consider the supplement data somehow a blackbox. The thing is we have to decide for either a \"not too good but super-robust\" or a \"really good with the risk of a total fail in the evaluation\" approach."
  },
  "source": "meta"
}