{
  "id": 249410,
  "title": "First place approach",
  "content": "<p>Hi all,</p>\n<p>thanks to the organizers for this challenging tabular competition and congrats to the top placed teams (it was extremely close in the end…)! I actually enjoyed the sprint character of this Data Cup very much (compared to typical competitions running 2 or 3 months)!</p>\n<p>I would like to share my approach which is based on the following public notebook:<br>\n<a href=\"https://www.kaggle.com/docxian/data-cup-round-2\" target=\"_blank\">https://www.kaggle.com/docxian/data-cup-round-2</a></p>\n<p>The best scoring approach was in the end a 60% / 40% blend of Version 15 (the blended output) and Version 5 (pure GBM). So it is basically a blend of two GBMs and a Random Forest model.</p>\n<p>Data Preparation steps (Version 15):</p>\n<ul>\n<li>Clean up date column (there were two different formats!)</li>\n<li>Clean up levels for atmosphere and statement</li>\n<li>Remove outliers for heat and rainfall (in training data)</li>\n<li>Extract hour, day, month, year and weekday from dt column</li>\n</ul>\n<p>The features year and flakes were not used they have no impact on the test set.<br>\nFurthermore weekday was NOT used. To be honest, I do not understand why this helps, but including it always led to a significantly worse result on the Public Leaderboard in spite of a much better Cross Validation. So the typcial \"trust the CV\"-rule did not really work here.</p>\n<p>Any ideas on that?</p>\n<p>Another somewhat strange thing is that Version 5 still includes the non-cleaned version of the dates, nevertheless it was the best performing individual model (public and private)!?</p>\n<p>Thanks again, keep up the good work and most important: have fun!</p>",
  "comments": [
    {
      "id": 1672204,
      "author_name": "Marília Prata",
      "author_url": "",
      "post_date": "2022-02-01T23:16:23.210000",
      "content": "<p>Delayed congrats Chris.  Well-done friend.</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "comment_count": 1,
  "messages": [
    {
      "id": 1672204,
      "postDate": "2022-02-01T23:16:23.210Z",
      "content": "<p>Delayed congrats Chris.  Well-done friend.</p>",
      "rawMarkdown": "Delayed congrats Chris.  Well-done friend.",
      "votes": 1
    },
    {
      "id": 1367952,
      "postDate": "2021-06-28T07:40:10.070Z",
      "content": "<p>Hi all,</p>\n<p>thanks to the organizers for this challenging tabular competition and congrats to the top placed teams (it was extremely close in the end…)! I actually enjoyed the sprint character of this Data Cup very much (compared to typical competitions running 2 or 3 months)!</p>\n<p>I would like to share my approach which is based on the following public notebook:<br>\n<a href=\"https://www.kaggle.com/docxian/data-cup-round-2\" target=\"_blank\">https://www.kaggle.com/docxian/data-cup-round-2</a></p>\n<p>The best scoring approach was in the end a 60% / 40% blend of Version 15 (the blended output) and Version 5 (pure GBM). So it is basically a blend of two GBMs and a Random Forest model.</p>\n<p>Data Preparation steps (Version 15):</p>\n<ul>\n<li>Clean up date column (there were two different formats!)</li>\n<li>Clean up levels for atmosphere and statement</li>\n<li>Remove outliers for heat and rainfall (in training data)</li>\n<li>Extract hour, day, month, year and weekday from dt column</li>\n</ul>\n<p>The features year and flakes were not used they have no impact on the test set.<br>\nFurthermore weekday was NOT used. To be honest, I do not understand why this helps, but including it always led to a significantly worse result on the Public Leaderboard in spite of a much better Cross Validation. So the typcial \"trust the CV\"-rule did not really work here.</p>\n<p>Any ideas on that?</p>\n<p>Another somewhat strange thing is that Version 5 still includes the non-cleaned version of the dates, nevertheless it was the best performing individual model (public and private)!?</p>\n<p>Thanks again, keep up the good work and most important: have fun!</p>",
      "rawMarkdown": "Hi all,\n\nthanks to the organizers for this challenging tabular competition and congrats to the top placed teams (it was extremely close in the end...)! I actually enjoyed the sprint character of this Data Cup very much (compared to typical competitions running 2 or 3 months)!\n\nI would like to share my approach which is based on the following public notebook:\nhttps://www.kaggle.com/docxian/data-cup-round-2\n\nThe best scoring approach was in the end a 60% / 40% blend of Version 15 (the blended output) and Version 5 (pure GBM). So it is basically a blend of two GBMs and a Random Forest model.\n\nData Preparation steps (Version 15):\n* Clean up date column (there were two different formats!)\n* Clean up levels for atmosphere and statement\n* Remove outliers for heat and rainfall (in training data)\n* Extract hour, day, month, year and weekday from dt column\n\nThe features year and flakes were not used they have no impact on the test set.\nFurthermore weekday was NOT used. To be honest, I do not understand why this helps, but including it always led to a significantly worse result on the Public Leaderboard in spite of a much better Cross Validation. So the typcial \"trust the CV\"-rule did not really work here.\n\nAny ideas on that?\n\nAnother somewhat strange thing is that Version 5 still includes the non-cleaned version of the dates, nevertheless it was the best performing individual model (public and private)!?\n\nThanks again, keep up the good work and most important: have fun!",
      "votes": 2
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-19T22:50:21Z",
  "competition_id": 29624,
  "competition_slug": "DATACUP-ROUND2",
  "team_id": 6997587,
  "solution_url": "/competitions/DATACUP-ROUND2/writeups/chris-x-first-place-approach"
}