{
  "id": 301012,
  "title": "Many are using LightGBM and variants for this competition, any reasons?",
  "url": "/competitions/g-research-crypto-forecasting/discussion/301012",
  "author_name": "",
  "post_date": "2022-01-15T12:28:54.454509100Z",
  "votes": 4,
  "comment_count": 8,
  "views": 0,
  "content": "<p>I have seen price sensitive datasets are expensive to train using these model algorithms:</p>\n<ul>\n<li>LightGBM</li>\n<li>XGBoost</li>\n<li>Catboost</li>\n</ul>\n<p>Of course its faster on the GPU but then the other model algorithms are even faster.</p>\n<p>Has anyone had better or competitive results by using MLP, or decision tree based or SVM like model algorithms? Neural Networks? Probabilistic, Bayesian models, Markov Models, or anything else in that space?</p>\n<p>What would be the advantages or disadvantages using one over the other in this competition?</p>\n<p>Lastly in theory classification algorithms could be applied to the datasets in this competition - has anyone tried this and has any good or not so good results?</p>\n<p>Would love to hear from you, as I have some experiences that I would love to share hereafter.</p>",
  "messages": [
    {
      "id": "1650927",
      "postDate": "01/15/2022 12:28:54",
      "content": "<p>I have seen price sensitive datasets are expensive to train using these model algorithms:</p>\n<ul>\n<li>LightGBM</li>\n<li>XGBoost</li>\n<li>Catboost</li>\n</ul>\n<p>Of course its faster on the GPU but then the other model algorithms are even faster.</p>\n<p>Has anyone had better or competitive results by using MLP, or decision tree based or SVM like model algorithms? Neural Networks? Probabilistic, Bayesian models, Markov Models, or anything else in that space?</p>\n<p>What would be the advantages or disadvantages using one over the other in this competition?</p>\n<p>Lastly in theory classification algorithms could be applied to the datasets in this competition - has anyone tried this and has any good or not so good results?</p>\n<p>Would love to hear from you, as I have some experiences that I would love to share hereafter.</p>",
      "rawMarkdown": "I have seen price sensitive datasets are expensive to train using these model algorithms:\n\n- LightGBM\n- XGBoost\n- Catboost\n\nOf course its faster on the GPU but then the other model algorithms are even faster.\n\nHas anyone had better or competitive results by using MLP, or decision tree based or SVM like model algorithms? Neural Networks? Probabilistic, Bayesian models, Markov Models, or anything else in that space?\n\nWhat would be the advantages or disadvantages using one over the other in this competition?\n\nLastly in theory classification algorithms could be applied to the datasets in this competition - has anyone tried this and has any good or not so good results?\n\nWould love to hear from you, as I have some experiences that I would love to share hereafter.",
      "votes": null
    },
    {
      "id": "1650938",
      "postDate": "01/15/2022 12:40:34",
      "content": "<p>Classification didn't perform as well when I did it.  But, it was a long time ago and my model has changed.  I feel like your not going to beat regression on a correlation matric task though.  Lightgbm is very fast so it should be comparable in speed to other algorithms.  SVM will probably be way to slow to train unless the training set you use is significantly cut down.  </p>",
      "rawMarkdown": "Classification didn't perform as well when I did it.  But, it was a long time ago and my model has changed.  I feel like your not going to beat regression on a correlation matric task though.  Lightgbm is very fast so it should be comparable in speed to other algorithms.  SVM will probably be way to slow to train unless the training set you use is significantly cut down.",
      "votes": null
    },
    {
      "id": "1650967",
      "postDate": "01/15/2022 13:10:16",
      "content": "<p>Thanks for the response, glad to see your shares about classification type of problem.</p>\n<p>I will have to do that training to see if that is the case, as I have had slow responses using LightGBM and other boosters with price data.</p>\n<p>Did you exclusively run them on the GPU? I have have to compare that benchmark and get back to you. </p>\n<p>I didn't say that my results are purely CPU based but I will run these on the GPU and see what transpires.</p>",
      "rawMarkdown": "Thanks for the response, glad to see your shares about classification type of problem.\n\nI will have to do that training to see if that is the case, as I have had slow responses using LightGBM and other boosters with price data.\n\nDid you exclusively run them on the GPU? I have have to compare that benchmark and get back to you. \n\nI didn't say that my results are purely CPU based but I will run these on the GPU and see what transpires.",
      "votes": null
    },
    {
      "id": "1651093",
      "postDate": "01/15/2022 14:18:23",
      "content": "<p>I'm not using a lightgbm model for the competition.  I was just saying that Lightgbm typically fits quickly.  I'm using a neural network with GPU because lightgbm performed poorly for me.</p>",
      "rawMarkdown": "I'm not using a lightgbm model for the competition.  I was just saying that Lightgbm typically fits quickly.  I'm using a neural network with GPU because lightgbm performed poorly for me.",
      "votes": null
    },
    {
      "id": "1651112",
      "postDate": "01/15/2022 14:38:12",
      "content": "<p>I was curious about the model algorithms for this competition - I have seen other mode algorithms perform better than the boosters (and faster), unless someone has some speed and model performance enhancements techniques then I would love to know.</p>\n<p>Is your NN solution public?</p>",
      "rawMarkdown": "I was curious about the model algorithms for this competition - I have seen other mode algorithms perform better than the boosters (and faster), unless someone has some speed and model performance enhancements techniques then I would love to know.\n\nIs your NN solution public?",
      "votes": null
    },
    {
      "id": "1651137",
      "postDate": "01/15/2022 15:03:22",
      "content": "<p>It is not, but there are some other public NN models.  </p>",
      "rawMarkdown": "It is not, but there are some other public NN models.",
      "votes": null
    },
    {
      "id": "1652412",
      "postDate": "01/16/2022 15:41:43",
      "content": "<p>As of now I have similar performance using GBDT / MLP on similar folds. The problem seems to be more about getting stable performance over time (i.e. dealing with non-stationnarity). The main difference are IMO:  MLP need GPU while lgbm do not. It's rather easy to customise the loss for MLP not for lgbm (require to derive gradient and hessian). However using the custom wieghted correlation loss doesn't really seems to improve performance over standard MAE. </p>",
      "rawMarkdown": "As of now I have similar performance using GBDT / MLP on similar folds. The problem seems to be more about getting stable performance over time (i.e. dealing with non-stationnarity). The main difference are IMO:  MLP need GPU while lgbm do not. It's rather easy to customise the loss for MLP not for lgbm (require to derive gradient and hessian). However using the custom wieghted correlation loss doesn't really seems to improve performance over standard MAE.",
      "votes": null
    },
    {
      "id": "1656236",
      "postDate": "01/19/2022 07:43:24",
      "content": "<p>because its a string algorithm that doesnt take  a toll on cpu processing</p>",
      "rawMarkdown": "because its a string algorithm that doesnt take  a toll on cpu processing",
      "votes": null
    },
    {
      "id": "1659732",
      "postDate": "01/22/2022 04:40:42",
      "content": "<p>I use LGBM only because of its fast speed…… XGBoost is really too slow for this competition. If you want to construct more features and try to ensemble models, you'd better use LGBM. XGBoost on GPU is also fast, but I can't appreciate its<br>\nprecision…</p>",
      "rawMarkdown": "I use LGBM only because of its fast speed...... XGBoost is really too slow for this competition. If you want to construct more features and try to ensemble models, you'd better use LGBM. XGBoost on GPU is also fast, but I can't appreciate its\nprecision...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1650938,
      "author_name": "petersk20",
      "author_url": "",
      "post_date": "01/15/2022 12:40:34",
      "content": "<p>Classification didn't perform as well when I did it.  But, it was a long time ago and my model has changed.  I feel like your not going to beat regression on a correlation matric task though.  Lightgbm is very fast so it should be comparable in speed to other algorithms.  SVM will probably be way to slow to train unless the training set you use is significantly cut down.  </p>",
      "votes": null,
      "replies": [
        {
          "id": 1650967,
          "author_name": "neomatrix369",
          "author_url": "",
          "post_date": "01/15/2022 13:10:16",
          "content": "<p>Thanks for the response, glad to see your shares about classification type of problem.</p>\n<p>I will have to do that training to see if that is the case, as I have had slow responses using LightGBM and other boosters with price data.</p>\n<p>Did you exclusively run them on the GPU? I have have to compare that benchmark and get back to you. </p>\n<p>I didn't say that my results are purely CPU based but I will run these on the GPU and see what transpires.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1651093,
          "author_name": "petersk20",
          "author_url": "",
          "post_date": "01/15/2022 14:18:23",
          "content": "<p>I'm not using a lightgbm model for the competition.  I was just saying that Lightgbm typically fits quickly.  I'm using a neural network with GPU because lightgbm performed poorly for me.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1651112,
          "author_name": "neomatrix369",
          "author_url": "",
          "post_date": "01/15/2022 14:38:12",
          "content": "<p>I was curious about the model algorithms for this competition - I have seen other mode algorithms perform better than the boosters (and faster), unless someone has some speed and model performance enhancements techniques then I would love to know.</p>\n<p>Is your NN solution public?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1651137,
          "author_name": "petersk20",
          "author_url": "",
          "post_date": "01/15/2022 15:03:22",
          "content": "<p>It is not, but there are some other public NN models.  </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1652412,
      "author_name": "lucasmorin",
      "author_url": "",
      "post_date": "01/16/2022 15:41:43",
      "content": "<p>As of now I have similar performance using GBDT / MLP on similar folds. The problem seems to be more about getting stable performance over time (i.e. dealing with non-stationnarity). The main difference are IMO:  MLP need GPU while lgbm do not. It's rather easy to customise the loss for MLP not for lgbm (require to derive gradient and hessian). However using the custom wieghted correlation loss doesn't really seems to improve performance over standard MAE. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1656236,
      "author_name": "muhammadammarjamshed",
      "author_url": "",
      "post_date": "01/19/2022 07:43:24",
      "content": "<p>because its a string algorithm that doesnt take  a toll on cpu processing</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1659732,
      "author_name": "noi031",
      "author_url": "",
      "post_date": "01/22/2022 04:40:42",
      "content": "<p>I use LGBM only because of its fast speed…… XGBoost is really too slow for this competition. If you want to construct more features and try to ensemble models, you'd better use LGBM. XGBoost on GPU is also fast, but I can't appreciate its<br>\nprecision…</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1650927": "I have seen price sensitive datasets are expensive to train using these model algorithms:\n\n- LightGBM\n- XGBoost\n- Catboost\n\nOf course its faster on the GPU but then the other model algorithms are even faster.\n\nHas anyone had better or competitive results by using MLP, or decision tree based or SVM like model algorithms? Neural Networks? Probabilistic, Bayesian models, Markov Models, or anything else in that space?\n\nWhat would be the advantages or disadvantages using one over the other in this competition?\n\nLastly in theory classification algorithms could be applied to the datasets in this competition - has anyone tried this and has any good or not so good results?\n\nWould love to hear from you, as I have some experiences that I would love to share hereafter.",
    "1650938": "Classification didn't perform as well when I did it.  But, it was a long time ago and my model has changed.  I feel like your not going to beat regression on a correlation matric task though.  Lightgbm is very fast so it should be comparable in speed to other algorithms.  SVM will probably be way to slow to train unless the training set you use is significantly cut down.",
    "1650967": "Thanks for the response, glad to see your shares about classification type of problem.\n\nI will have to do that training to see if that is the case, as I have had slow responses using LightGBM and other boosters with price data.\n\nDid you exclusively run them on the GPU? I have have to compare that benchmark and get back to you. \n\nI didn't say that my results are purely CPU based but I will run these on the GPU and see what transpires.",
    "1651093": "I'm not using a lightgbm model for the competition.  I was just saying that Lightgbm typically fits quickly.  I'm using a neural network with GPU because lightgbm performed poorly for me.",
    "1651112": "I was curious about the model algorithms for this competition - I have seen other mode algorithms perform better than the boosters (and faster), unless someone has some speed and model performance enhancements techniques then I would love to know.\n\nIs your NN solution public?",
    "1651137": "It is not, but there are some other public NN models.",
    "1652412": "As of now I have similar performance using GBDT / MLP on similar folds. The problem seems to be more about getting stable performance over time (i.e. dealing with non-stationnarity). The main difference are IMO:  MLP need GPU while lgbm do not. It's rather easy to customise the loss for MLP not for lgbm (require to derive gradient and hessian). However using the custom wieghted correlation loss doesn't really seems to improve performance over standard MAE.",
    "1656236": "because its a string algorithm that doesnt take  a toll on cpu processing",
    "1659732": "I use LGBM only because of its fast speed...... XGBoost is really too slow for this competition. If you want to construct more features and try to ensemble models, you'd better use LGBM. XGBoost on GPU is also fast, but I can't appreciate its\nprecision..."
  },
  "source": "meta"
}