{
  "id": 288216,
  "title": "is being able to recalculate the target a 'must' for training?",
  "url": "/competitions/g-research-crypto-forecasting/discussion/288216",
  "author_name": "",
  "post_date": "2021-11-17T07:26:31.582013100Z",
  "votes": 4,
  "comment_count": 5,
  "views": 0,
  "content": "<p>If the host clearly confirms in the pinned discussion <a href=\"https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/286779\" target=\"_blank\">in this link</a>, that, target will be provided in all supplemental and test data, is there an explicit first order value of calculating the target? Thanks.</p>",
  "messages": [
    {
      "id": "1585226",
      "postDate": "11/17/2021 07:26:31",
      "content": "<p>If the host clearly confirms in the pinned discussion <a href=\"https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/286779\" target=\"_blank\">in this link</a>, that, target will be provided in all supplemental and test data, is there an explicit first order value of calculating the target? Thanks.</p>",
      "rawMarkdown": "If the host clearly confirms in the pinned discussion [in this link](https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/286779), that, target will be provided in all supplemental and test data, is there an explicit first order value of calculating the target? Thanks.",
      "votes": null
    },
    {
      "id": "1585252",
      "postDate": "11/17/2021 07:55:52",
      "content": "<blockquote>\n  <p>target will be provided in all supplemental and <strong>test data</strong></p>\n</blockquote>\n<p>What do you mean by test data here? I haven't seen them say anywhere that Targets would be available during the test (Private LB) period.</p>",
      "rawMarkdown": "> target will be provided in all supplemental and **test data**\n\nWhat do you mean by test data here? I haven't seen them say anywhere that Targets would be available during the test (Private LB) period.",
      "votes": null
    },
    {
      "id": "1585320",
      "postDate": "11/17/2021 08:31:40",
      "content": "<p>Sorry for causing misunderstanding, my bad.<br>\nI meant the supplemantal data in the period upto submission and additional data until testtime, so that the data is contiguous. </p>\n<p>And my question was (to rephrase the topic question), if being able to re-calculate the target is a 'must' or 'a less stringent condition', because for baseline models (baseline booster algorithms and also NN's) I may get away using 'the target' in train (and as will in supplemental-) data 'as is', without building features based on parameters used to calculate the target?</p>",
      "rawMarkdown": "Sorry for causing misunderstanding, my bad.\nI meant the supplemantal data in the period upto submission and additional data until testtime, so that the data is contiguous. \n\nAnd my question was (to rephrase the topic question), if being able to re-calculate the target is a 'must' or 'a less stringent condition', because for baseline models (baseline booster algorithms and also NN's) I may get away using 'the target' in train (and as will in supplemental-) data 'as is', without building features based on parameters used to calculate the target?",
      "votes": null
    },
    {
      "id": "1585570",
      "postDate": "11/17/2021 10:52:58",
      "content": "<p>The target is rather unusual and involve some specific things. Notably it involves some 'market' and beta with respect to this market. Understanding how the target is built might help for feature engineering: creating agggregate features over this 'market', building betas as features…</p>",
      "rawMarkdown": "The target is rather unusual and involve some specific things. Notably it involves some 'market' and beta with respect to this market. Understanding how the target is built might help for feature engineering: creating agggregate features over this 'market', building betas as features...",
      "votes": null
    },
    {
      "id": "1585829",
      "postDate": "11/17/2021 15:23:09",
      "content": "<p>Obviously you can create a model without knowing how the target was calculated, but how your model can predict with good accuracy if the target include information you are not using in your features?</p>\n<p>Example, target is normalized with values that ranges 3750 minutes span, if you are not giving that information to your model the score is not going to be as good as possible.</p>\n<p>And of course, this is Kaggle and we love a good mathematical puzzle.</p>",
      "rawMarkdown": "Obviously you can create a model without knowing how the target was calculated, but how your model can predict with good accuracy if the target include information you are not using in your features?\n\nExample, target is normalized with values that ranges 3750 minutes span, if you are not giving that information to your model the score is not going to be as good as possible.\n\nAnd of course, this is Kaggle and we love a good mathematical puzzle.",
      "votes": null
    },
    {
      "id": "1586756",
      "postDate": "11/18/2021 08:44:56",
      "content": "<p>I agree, we can assume our model is able to predict future returns, but it has no way of learning how to subtract the extra term from these returns to get the targets (unless we deal with a 3750 min long time series). On a 3-month time scale, beta is far from being constant. Instead we can just supply the last calculated beta to the model as another parameter. </p>",
      "rawMarkdown": "I agree, we can assume our model is able to predict future returns, but it has no way of learning how to subtract the extra term from these returns to get the targets (unless we deal with a 3750 min long time series). On a 3-month time scale, beta is far from being constant. Instead we can just supply the last calculated beta to the model as another parameter.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1585252,
      "author_name": "brandenkmurray",
      "author_url": "",
      "post_date": "11/17/2021 07:55:52",
      "content": "<blockquote>\n  <p>target will be provided in all supplemental and <strong>test data</strong></p>\n</blockquote>\n<p>What do you mean by test data here? I haven't seen them say anywhere that Targets would be available during the test (Private LB) period.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1585320,
          "author_name": "abdulkadirguner",
          "author_url": "",
          "post_date": "11/17/2021 08:31:40",
          "content": "<p>Sorry for causing misunderstanding, my bad.<br>\nI meant the supplemantal data in the period upto submission and additional data until testtime, so that the data is contiguous. </p>\n<p>And my question was (to rephrase the topic question), if being able to re-calculate the target is a 'must' or 'a less stringent condition', because for baseline models (baseline booster algorithms and also NN's) I may get away using 'the target' in train (and as will in supplemental-) data 'as is', without building features based on parameters used to calculate the target?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1585570,
      "author_name": "lucasmorin",
      "author_url": "",
      "post_date": "11/17/2021 10:52:58",
      "content": "<p>The target is rather unusual and involve some specific things. Notably it involves some 'market' and beta with respect to this market. Understanding how the target is built might help for feature engineering: creating agggregate features over this 'market', building betas as features…</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1585829,
      "author_name": "agenlu",
      "author_url": "",
      "post_date": "11/17/2021 15:23:09",
      "content": "<p>Obviously you can create a model without knowing how the target was calculated, but how your model can predict with good accuracy if the target include information you are not using in your features?</p>\n<p>Example, target is normalized with values that ranges 3750 minutes span, if you are not giving that information to your model the score is not going to be as good as possible.</p>\n<p>And of course, this is Kaggle and we love a good mathematical puzzle.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1586756,
          "author_name": "sergears",
          "author_url": "",
          "post_date": "11/18/2021 08:44:56",
          "content": "<p>I agree, we can assume our model is able to predict future returns, but it has no way of learning how to subtract the extra term from these returns to get the targets (unless we deal with a 3750 min long time series). On a 3-month time scale, beta is far from being constant. Instead we can just supply the last calculated beta to the model as another parameter. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1585226": "If the host clearly confirms in the pinned discussion [in this link](https://www.kaggle.com/c/g-research-crypto-forecasting/discussion/286779), that, target will be provided in all supplemental and test data, is there an explicit first order value of calculating the target? Thanks.",
    "1585252": "> target will be provided in all supplemental and **test data**\n\nWhat do you mean by test data here? I haven't seen them say anywhere that Targets would be available during the test (Private LB) period.",
    "1585320": "Sorry for causing misunderstanding, my bad.\nI meant the supplemantal data in the period upto submission and additional data until testtime, so that the data is contiguous. \n\nAnd my question was (to rephrase the topic question), if being able to re-calculate the target is a 'must' or 'a less stringent condition', because for baseline models (baseline booster algorithms and also NN's) I may get away using 'the target' in train (and as will in supplemental-) data 'as is', without building features based on parameters used to calculate the target?",
    "1585570": "The target is rather unusual and involve some specific things. Notably it involves some 'market' and beta with respect to this market. Understanding how the target is built might help for feature engineering: creating agggregate features over this 'market', building betas as features...",
    "1585829": "Obviously you can create a model without knowing how the target was calculated, but how your model can predict with good accuracy if the target include information you are not using in your features?\n\nExample, target is normalized with values that ranges 3750 minutes span, if you are not giving that information to your model the score is not going to be as good as possible.\n\nAnd of course, this is Kaggle and we love a good mathematical puzzle.",
    "1586756": "I agree, we can assume our model is able to predict future returns, but it has no way of learning how to subtract the extra term from these returns to get the targets (unless we deal with a 3750 min long time series). On a 3-month time scale, beta is far from being constant. Instead we can just supply the last calculated beta to the model as another parameter."
  },
  "source": "meta"
}