{
  "id": 285957,
  "title": "Updating features for predictions",
  "url": "/competitions/g-research-crypto-forecasting/discussion/285957",
  "author_name": "",
  "post_date": "2021-11-07T08:47:56.031140Z",
  "votes": 8,
  "comment_count": 2,
  "views": 0,
  "content": "<p>What approaches are you taking for updating features for predicting on the test data? (Especially given the constraints of the API)?</p>\n<p>e.g. updating lag/time series features, moving averages etc', without recalculating the whole historical context. </p>",
  "messages": [
    {
      "id": "1574109",
      "postDate": "11/07/2021 08:47:56",
      "content": "<p>What approaches are you taking for updating features for predicting on the test data? (Especially given the constraints of the API)?</p>\n<p>e.g. updating lag/time series features, moving averages etc', without recalculating the whole historical context. </p>",
      "rawMarkdown": "What approaches are you taking for updating features for predicting on the test data? (Especially given the constraints of the API)?\n\ne.g. updating lag/time series features, moving averages etc', without recalculating the whole historical context.",
      "votes": null
    },
    {
      "id": "1589096",
      "postDate": "11/19/2021 23:22:13",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/danofer\" target=\"_blank\">@danofer</a>, <br>\nI realize this issue trying to build some Lag features for my model, basically because the API iterator release1 min. of data at the time for all the assets and because you need to predict something before you can see the next round of data, probably is hard to use Lag features. <br>\nI thought maybe to store them as they are released and update my predictions as more data is accumulated; the issue is that I will never know when to start updating because the iterator can run for 100 cycles or ten cycles? And eventually, my implementation will fail.<br>\nAny ideas are appreciated.</p>",
      "rawMarkdown": "Hello @danofer, \nI realize this issue trying to build some Lag features for my model, basically because the API iterator release1 min. of data at the time for all the assets and because you need to predict something before you can see the next round of data, probably is hard to use Lag features. \nI thought maybe to store them as they are released and update my predictions as more data is accumulated; the issue is that I will never know when to start updating because the iterator can run for 100 cycles or ten cycles? And eventually, my implementation will fail.\nAny ideas are appreciated.",
      "votes": null
    },
    {
      "id": "1597790",
      "postDate": "11/27/2021 22:32:55",
      "content": "<p>I think this is bad idea<br>\nReasons<br>\n1)Lags did not provide any information, because crypto for example very change based on news, volume change, <br>\n2)On the 2sigma competition one guy tried to create solutions based on different lags and was on 50 best or Public LB, but after new data, he goes down to 1000. :) </p>",
      "rawMarkdown": "I think this is bad idea\nReasons\n1)Lags did not provide any information, because crypto for example very change based on news, volume change, \n2)On the 2sigma competition one guy tried to create solutions based on different lags and was on 50 best or Public LB, but after new data, he goes down to 1000. :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1589096,
      "author_name": "cv13j0",
      "author_url": "",
      "post_date": "11/19/2021 23:22:13",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/danofer\" target=\"_blank\">@danofer</a>, <br>\nI realize this issue trying to build some Lag features for my model, basically because the API iterator release1 min. of data at the time for all the assets and because you need to predict something before you can see the next round of data, probably is hard to use Lag features. <br>\nI thought maybe to store them as they are released and update my predictions as more data is accumulated; the issue is that I will never know when to start updating because the iterator can run for 100 cycles or ten cycles? And eventually, my implementation will fail.<br>\nAny ideas are appreciated.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1597790,
          "author_name": "vladimircape",
          "author_url": "",
          "post_date": "11/27/2021 22:32:55",
          "content": "<p>I think this is bad idea<br>\nReasons<br>\n1)Lags did not provide any information, because crypto for example very change based on news, volume change, <br>\n2)On the 2sigma competition one guy tried to create solutions based on different lags and was on 50 best or Public LB, but after new data, he goes down to 1000. :) </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1574109": "What approaches are you taking for updating features for predicting on the test data? (Especially given the constraints of the API)?\n\ne.g. updating lag/time series features, moving averages etc', without recalculating the whole historical context.",
    "1589096": "Hello @danofer, \nI realize this issue trying to build some Lag features for my model, basically because the API iterator release1 min. of data at the time for all the assets and because you need to predict something before you can see the next round of data, probably is hard to use Lag features. \nI thought maybe to store them as they are released and update my predictions as more data is accumulated; the issue is that I will never know when to start updating because the iterator can run for 100 cycles or ten cycles? And eventually, my implementation will fail.\nAny ideas are appreciated.",
    "1597790": "I think this is bad idea\nReasons\n1)Lags did not provide any information, because crypto for example very change based on news, volume change, \n2)On the 2sigma competition one guy tried to create solutions based on different lags and was on 50 best or Public LB, but after new data, he goes down to 1000. :)"
  },
  "source": "meta"
}