{
  "id": 301946,
  "title": "GridSearchCV does not seem to work with LightGBM",
  "url": "/competitions/g-research-crypto-forecasting/discussion/301946",
  "author_name": "",
  "post_date": "2022-01-20T07:06:56.504427400Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello! I've tried several times LightGBM with default parameters and with GridSearchCV from scikit-learn and it turns out that default parameters always perform better, even if default parameters are included in grid search. As a metric in grid search, I use a custom function that calculates Pearson correlation. Am I missing something?</p>",
  "messages": [
    {
      "id": "1657426",
      "postDate": "01/20/2022 07:06:56",
      "content": "<p>Hello! I've tried several times LightGBM with default parameters and with GridSearchCV from scikit-learn and it turns out that default parameters always perform better, even if default parameters are included in grid search. As a metric in grid search, I use a custom function that calculates Pearson correlation. Am I missing something?</p>",
      "rawMarkdown": "Hello! I've tried several times LightGBM with default parameters and with GridSearchCV from scikit-learn and it turns out that default parameters always perform better, even if default parameters are included in grid search. As a metric in grid search, I use a custom function that calculates Pearson correlation. Am I missing something?",
      "votes": null
    },
    {
      "id": "1658287",
      "postDate": "01/20/2022 22:29:12",
      "content": "<p>I have had that happen to me. I don't think it's bad, but it makes me think its possible that all my feature selection could have been biased towards these parameters or something like that. Or the defaults are just really good in general. </p>",
      "rawMarkdown": "I have had that happen to me. I don't think it's bad, but it makes me think its possible that all my feature selection could have been biased towards these parameters or something like that. Or the defaults are just really good in general.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1658287,
      "author_name": "chrisrichardmiles",
      "author_url": "",
      "post_date": "01/20/2022 22:29:12",
      "content": "<p>I have had that happen to me. I don't think it's bad, but it makes me think its possible that all my feature selection could have been biased towards these parameters or something like that. Or the defaults are just really good in general. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1657426": "Hello! I've tried several times LightGBM with default parameters and with GridSearchCV from scikit-learn and it turns out that default parameters always perform better, even if default parameters are included in grid search. As a metric in grid search, I use a custom function that calculates Pearson correlation. Am I missing something?",
    "1658287": "I have had that happen to me. I don't think it's bad, but it makes me think its possible that all my feature selection could have been biased towards these parameters or something like that. Or the defaults are just really good in general."
  },
  "source": "meta"
}