{
  "id": 680578,
  "title": "Readmission rate vs prediction distribution",
  "url": "/competitions/ai-emergency-medicine-turkiye/discussion/680578",
  "author_name": "Cem Karisli",
  "post_date": "2026-03-09T13:20:06.286000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n<p>According to the competition description, the average rate of 30-day readmission is around 13%. However, when I look at the predictions generated by my models, the mean of the predicted probabilities in the submission files is usually around 0.40–0.50.</p>\n<p>Despite this, the model performs well on AUC and currently ranks well on the leaderboard.</p>\n<p>This made me wonder a few things:</p>\n<p>Is this behaviour common for others as well?</p>\n<p>Could this be related to using class balancing techniques (e.g. CatBoost auto_class_weights=\"Balanced\" or similar)?</p>\n<p>Or is it simply expected since AUC evaluates ranking rather than probability calibration?</p>\n<p>In other words, should we worry if the predicted probability distribution is far from the actual event rate, as long as AUC is strong?</p>\n<p>Curious to hear how others are approaching this.</p>\n<p>Thanks!</p>",
  "messages": [
    {
      "id": 3418896,
      "postDate": "2026-03-09T13:20:06.287Z",
      "content": "<p>Hi everyone,</p>\n<p>According to the competition description, the average rate of 30-day readmission is around 13%. However, when I look at the predictions generated by my models, the mean of the predicted probabilities in the submission files is usually around 0.40–0.50.</p>\n<p>Despite this, the model performs well on AUC and currently ranks well on the leaderboard.</p>\n<p>This made me wonder a few things:</p>\n<p>Is this behaviour common for others as well?</p>\n<p>Could this be related to using class balancing techniques (e.g. CatBoost auto_class_weights=\"Balanced\" or similar)?</p>\n<p>Or is it simply expected since AUC evaluates ranking rather than probability calibration?</p>\n<p>In other words, should we worry if the predicted probability distribution is far from the actual event rate, as long as AUC is strong?</p>\n<p>Curious to hear how others are approaching this.</p>\n<p>Thanks!</p>",
      "rawMarkdown": "Hi everyone,\n\nAccording to the competition description, the average rate of 30-day readmission is around 13%. However, when I look at the predictions generated by my models, the mean of the predicted probabilities in the submission files is usually around 0.40–0.50.\n\nDespite this, the model performs well on AUC and currently ranks well on the leaderboard.\n\nThis made me wonder a few things:\n\nIs this behaviour common for others as well?\n\nCould this be related to using class balancing techniques (e.g. CatBoost auto_class_weights=\"Balanced\" or similar)?\n\nOr is it simply expected since AUC evaluates ranking rather than probability calibration?\n\nIn other words, should we worry if the predicted probability distribution is far from the actual event rate, as long as AUC is strong?\n\nCurious to hear how others are approaching this.\n\nThanks!"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3418896": "Hi everyone,\n\nAccording to the competition description, the average rate of 30-day readmission is around 13%. However, when I look at the predictions generated by my models, the mean of the predicted probabilities in the submission files is usually around 0.40–0.50.\n\nDespite this, the model performs well on AUC and currently ranks well on the leaderboard.\n\nThis made me wonder a few things:\n\nIs this behaviour common for others as well?\n\nCould this be related to using class balancing techniques (e.g. CatBoost auto_class_weights=\"Balanced\" or similar)?\n\nOr is it simply expected since AUC evaluates ranking rather than probability calibration?\n\nIn other words, should we worry if the predicted probability distribution is far from the actual event rate, as long as AUC is strong?\n\nCurious to hear how others are approaching this.\n\nThanks!"
  }
}