{
  "id": 323219,
  "title": "Submission Stratgy",
  "url": "/competitions/g-research-crypto-forecasting/discussion/323219",
  "author_name": "",
  "post_date": "2022-05-05T09:58:22.063287900Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>First of all, congratulations to all the winners. Also, I feel the pain of the competitors who put huge effort and still got unlucky because of the submission runtime or some other unforeseen glitch. </p>\n<p>In this post, I will share my submission strategy which I thought would be helpful but was not much eventually. I observed that Bitcoin and Ethereum forecasts are always highly correlated, so in one submission, I just took the negative values of their predictions, and in another one, I kept them as they are. I felt that in case, my model does not work for these two highly weighted coins, the other submission would mitigate that issue. However, this strategy works well if your model does not predict almost ~0.00xx absolute correlation for these two coins. Unfortunately, that's what happened as my two models almost had similar performances eventually and both submissions had close final scores. </p>\n<p>Just in case, you are interested, I only used one single feature from the data: \"close\" and engineered some time-series features from that. Also, had different features for different coins and used simple OLS regression.</p>\n<p>Cheers</p>",
  "messages": [
    {
      "id": "1778444",
      "postDate": "05/05/2022 09:58:22",
      "content": "<p>First of all, congratulations to all the winners. Also, I feel the pain of the competitors who put huge effort and still got unlucky because of the submission runtime or some other unforeseen glitch. </p>\n<p>In this post, I will share my submission strategy which I thought would be helpful but was not much eventually. I observed that Bitcoin and Ethereum forecasts are always highly correlated, so in one submission, I just took the negative values of their predictions, and in another one, I kept them as they are. I felt that in case, my model does not work for these two highly weighted coins, the other submission would mitigate that issue. However, this strategy works well if your model does not predict almost ~0.00xx absolute correlation for these two coins. Unfortunately, that's what happened as my two models almost had similar performances eventually and both submissions had close final scores. </p>\n<p>Just in case, you are interested, I only used one single feature from the data: \"close\" and engineered some time-series features from that. Also, had different features for different coins and used simple OLS regression.</p>\n<p>Cheers</p>",
      "rawMarkdown": "First of all, congratulations to all the winners. Also, I feel the pain of the competitors who put huge effort and still got unlucky because of the submission runtime or some other unforeseen glitch. \n\nIn this post, I will share my submission strategy which I thought would be helpful but was not much eventually. I observed that Bitcoin and Ethereum forecasts are always highly correlated, so in one submission, I just took the negative values of their predictions, and in another one, I kept them as they are. I felt that in case, my model does not work for these two highly weighted coins, the other submission would mitigate that issue. However, this strategy works well if your model does not predict almost ~0.00xx absolute correlation for these two coins. Unfortunately, that's what happened as my two models almost had similar performances eventually and both submissions had close final scores. \n\nJust in case, you are interested, I only used one single feature from the data: \"close\" and engineered some time-series features from that. Also, had different features for different coins and used simple OLS regression.\n\nCheers",
      "votes": null
    },
    {
      "id": "1779207",
      "postDate": "05/06/2022 07:41:01",
      "content": "<p>Your submission strategy is interesting and I can see how it would work well if the correlation between the two coins was not so close to zero. In this case, it sounds like your two submissions cancelling each other out.</p>\n<p>Still, thanks for sharing your thoughts and I hope it helps someone else out there.</p>",
      "rawMarkdown": "Your submission strategy is interesting and I can see how it would work well if the correlation between the two coins was not so close to zero. In this case, it sounds like your two submissions cancelling each other out.\n\nStill, thanks for sharing your thoughts and I hope it helps someone else out there.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1779207,
      "author_name": "",
      "author_url": "",
      "post_date": "05/06/2022 07:41:01",
      "content": "<p>Your submission strategy is interesting and I can see how it would work well if the correlation between the two coins was not so close to zero. In this case, it sounds like your two submissions cancelling each other out.</p>\n<p>Still, thanks for sharing your thoughts and I hope it helps someone else out there.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1778444": "First of all, congratulations to all the winners. Also, I feel the pain of the competitors who put huge effort and still got unlucky because of the submission runtime or some other unforeseen glitch. \n\nIn this post, I will share my submission strategy which I thought would be helpful but was not much eventually. I observed that Bitcoin and Ethereum forecasts are always highly correlated, so in one submission, I just took the negative values of their predictions, and in another one, I kept them as they are. I felt that in case, my model does not work for these two highly weighted coins, the other submission would mitigate that issue. However, this strategy works well if your model does not predict almost ~0.00xx absolute correlation for these two coins. Unfortunately, that's what happened as my two models almost had similar performances eventually and both submissions had close final scores. \n\nJust in case, you are interested, I only used one single feature from the data: \"close\" and engineered some time-series features from that. Also, had different features for different coins and used simple OLS regression.\n\nCheers",
    "1779207": "Your submission strategy is interesting and I can see how it would work well if the correlation between the two coins was not so close to zero. In this case, it sounds like your two submissions cancelling each other out.\n\nStill, thanks for sharing your thoughts and I hope it helps someone else out there."
  },
  "source": "meta"
}