{
  "id": 613719,
  "title": "Why Are There Negative Scores on the Leaderboard?",
  "url": "/competitions/csiro-biomass/discussion/613719",
  "author_name": "",
  "post_date": "2025-10-29T07:31:59.704403500Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Based on the evaluation metric (the coefficient of determination), negative scores shouldn’t be possible. Does this mean the issue comes from how the leaderboard calculates or normalizes the scores?</p>",
  "messages": [
    {
      "id": "3308354",
      "postDate": "10/29/2025 07:31:59",
      "content": "<p>Based on the evaluation metric (the coefficient of determination), negative scores shouldn’t be possible. Does this mean the issue comes from how the leaderboard calculates or normalizes the scores?</p>",
      "rawMarkdown": "Based on the evaluation metric (the coefficient of determination), negative scores shouldn’t be possible. Does this mean the issue comes from how the leaderboard calculates or normalizes the scores?",
      "votes": null
    },
    {
      "id": "3308361",
      "postDate": "10/29/2025 07:58:55",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/gunianxue\" target=\"_blank\">@gunianxue</a>!\nNegative R² values can occur and don’t indicate a leaderboard or normalization issue. They mean that the model’s predictions are performing worse than simply predicting the mean of the target values. </p>",
      "rawMarkdown": "Hi @gunianxue!\nNegative R² values can occur and don’t indicate a leaderboard or normalization issue. They mean that the model’s predictions are performing worse than simply predicting the mean of the target values.",
      "votes": null
    },
    {
      "id": "3308362",
      "postDate": "10/29/2025 08:01:50",
      "content": "<p>It's actually possible to get a negative R^2 score. Recall that \n$$\n    R^2 = 1 - \\frac{SS_{res}}{SS_{tot}}\n$$\nso if the model's MSE (SS_res) is greater than the variance of your target variable (SS_tot) (which is also the optimal MSE for a mean prediction) (which can happen if you for example double or halve your predictions accidentally, or there are some really big outliers disrupting the training or the evaluation score, or other reasons…), one can actually get a negative R^2 score.</p>",
      "rawMarkdown": "It's actually possible to get a negative R^2 score. Recall that \n$$\n    R^2 = 1 - \\frac{SS_{res}}{SS_{tot}}\n$$\nso if the model's MSE (SS_res) is greater than the variance of your target variable (SS_tot) (which is also the optimal MSE for a mean prediction) (which can happen if you for example double or halve your predictions accidentally, or there are some really big outliers disrupting the training or the evaluation score, or other reasons...), one can actually get a negative R^2 score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3308361,
      "author_name": "martynaplomecka",
      "author_url": "",
      "post_date": "10/29/2025 07:58:55",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/gunianxue\" target=\"_blank\">@gunianxue</a>!\nNegative R² values can occur and don’t indicate a leaderboard or normalization issue. They mean that the model’s predictions are performing worse than simply predicting the mean of the target values. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3308362,
      "author_name": "carsoncheng",
      "author_url": "",
      "post_date": "10/29/2025 08:01:50",
      "content": "<p>It's actually possible to get a negative R^2 score. Recall that \n$$\n    R^2 = 1 - \\frac{SS_{res}}{SS_{tot}}\n$$\nso if the model's MSE (SS_res) is greater than the variance of your target variable (SS_tot) (which is also the optimal MSE for a mean prediction) (which can happen if you for example double or halve your predictions accidentally, or there are some really big outliers disrupting the training or the evaluation score, or other reasons…), one can actually get a negative R^2 score.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3308354": "Based on the evaluation metric (the coefficient of determination), negative scores shouldn’t be possible. Does this mean the issue comes from how the leaderboard calculates or normalizes the scores?",
    "3308361": "Hi @gunianxue!\nNegative R² values can occur and don’t indicate a leaderboard or normalization issue. They mean that the model’s predictions are performing worse than simply predicting the mean of the target values.",
    "3308362": "It's actually possible to get a negative R^2 score. Recall that \n$$\n    R^2 = 1 - \\frac{SS_{res}}{SS_{tot}}\n$$\nso if the model's MSE (SS_res) is greater than the variance of your target variable (SS_tot) (which is also the optimal MSE for a mean prediction) (which can happen if you for example double or halve your predictions accidentally, or there are some really big outliers disrupting the training or the evaluation score, or other reasons...), one can actually get a negative R^2 score."
  },
  "source": "meta"
}