{
  "id": 589020,
  "title": "What happens if predictions lie outside of confidence intervals?",
  "url": "/competitions/ariel-data-challenge-2025/discussion/589020",
  "author_name": "0STG0T",
  "post_date": "2025-07-09T18:09:12.032000",
  "votes": 0,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello, I am trying out in this competition and lately was getting some zero score submissions, but on crossval i got scores near 0.35 also using OOF. How does the metric compute when the predicted wavelengths intensities lie outside of confidence intervals?</p>",
  "messages": [
    {
      "id": 3245770,
      "postDate": "2025-07-09T20:36:05.127Z",
      "content": "<p>You mean your mu_pred is outside y_true ± sigma_pred? It follows the equation in the overview page. You will have quadratically increasing penalty from the Mean Square Error term in the competition metric GLL: [(mu_pred - y_true) / sigma_pred] ** 2.</p>\n<p>Wavelength dependence is difficult, so it's probably good to start from wavelength-independent prediction (same for all 283 columns) and constant sigma like 1e-3. </p>",
      "rawMarkdown": "You mean your mu_pred is outside y_true ± sigma_pred? It follows the equation in the overview page. You will have quadratically increasing penalty from the Mean Square Error term in the competition metric GLL: [(mu_pred - y_true) / sigma_pred] ** 2.\n\nWavelength dependence is difficult, so it's probably good to start from wavelength-independent prediction (same for all 283 columns) and constant sigma like 1e-3. ",
      "votes": 2,
      "replies": [
        {
          "id": 3245780,
          "postDate": "2025-07-09T20:52:59.400Z",
          "content": "<p>Thank you for your reply, it was helpful! I have mistaken in my statement, I meant y_true is outside mu_pred ± sigma_pred. Now i get it, in the MSE term we are getting ratio of actual error interval to the error interval that we bet on</p>",
          "rawMarkdown": "Thank you for your reply, it was helpful! I have mistaken in my statement, I meant y_true is outside mu_pred ± sigma_pred. Now i get it, in the MSE term we are getting ratio of actual error interval to the error interval that we bet on",
          "votes": 1
        }
      ]
    },
    {
      "id": 3245717,
      "postDate": "2025-07-09T18:09:12.033Z",
      "content": "<p>Hello, I am trying out in this competition and lately was getting some zero score submissions, but on crossval i got scores near 0.35 also using OOF. How does the metric compute when the predicted wavelengths intensities lie outside of confidence intervals?</p>",
      "rawMarkdown": "Hello, I am trying out in this competition and lately was getting some zero score submissions, but on crossval i got scores near 0.35 also using OOF. How does the metric compute when the predicted wavelengths intensities lie outside of confidence intervals?"
    }
  ],
  "comments": [
    {
      "id": 3245770,
      "author_name": "🐢 Jun Koda",
      "author_url": "",
      "post_date": "2025-07-09T20:36:05.127000",
      "content": "<p>You mean your mu_pred is outside y_true ± sigma_pred? It follows the equation in the overview page. You will have quadratically increasing penalty from the Mean Square Error term in the competition metric GLL: [(mu_pred - y_true) / sigma_pred] ** 2.</p>\n<p>Wavelength dependence is difficult, so it's probably good to start from wavelength-independent prediction (same for all 283 columns) and constant sigma like 1e-3. </p>",
      "votes": 2,
      "replies": [
        {
          "id": 3245780,
          "author_name": "0STG0T",
          "author_url": "",
          "post_date": "2025-07-09T20:52:59.400000",
          "content": "<p>Thank you for your reply, it was helpful! I have mistaken in my statement, I meant y_true is outside mu_pred ± sigma_pred. Now i get it, in the MSE term we are getting ratio of actual error interval to the error interval that we bet on</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3245770": "You mean your mu_pred is outside y_true ± sigma_pred? It follows the equation in the overview page. You will have quadratically increasing penalty from the Mean Square Error term in the competition metric GLL: [(mu_pred - y_true) / sigma_pred] ** 2.\n\nWavelength dependence is difficult, so it's probably good to start from wavelength-independent prediction (same for all 283 columns) and constant sigma like 1e-3. ",
    "3245717": "Hello, I am trying out in this competition and lately was getting some zero score submissions, but on crossval i got scores near 0.35 also using OOF. How does the metric compute when the predicted wavelengths intensities lie outside of confidence intervals?"
  }
}