{
  "id": 118197,
  "title": "Model Type",
  "url": "/competitions/ashrae-energy-prediction/discussion/118197",
  "author_name": "",
  "post_date": "2019-11-20T01:38:31.319998100Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello!</p>\n\n<p>I noticed a number of notebooks using LightGBM to perform modeling.\nIs there any particular reason for this?\nI did a little bit of research and found that VAR and VARMAX can be used to  model time-series forecasting but did not find any notebooks related to these. \nAm I missing anything here?\nAny help would be appreciated. Thanks!</p>",
  "messages": [
    {
      "id": "677276",
      "postDate": "11/20/2019 01:38:31",
      "content": "<p>Hello!</p>\n\n<p>I noticed a number of notebooks using LightGBM to perform modeling.\nIs there any particular reason for this?\nI did a little bit of research and found that VAR and VARMAX can be used to  model time-series forecasting but did not find any notebooks related to these. \nAm I missing anything here?\nAny help would be appreciated. Thanks!</p>",
      "rawMarkdown": "Hello!\n\nI noticed a number of notebooks using LightGBM to perform modeling.\nIs there any particular reason for this?\nI did a little bit of research and found that VAR and VARMAX can be used to  model time-series forecasting but did not find any notebooks related to these. \nAm I missing anything here?\nAny help would be appreciated. Thanks!",
      "votes": null
    },
    {
      "id": "677384",
      "postDate": "11/20/2019 04:47:55",
      "content": "<p>Hi ,  In this competition we have to forecast for 2 years in the future and the models that you mentioned doesn't work very well in cases where it has to do a multistep forecast for so many steps </p>",
      "rawMarkdown": "Hi ,  In this competition we have to forecast for 2 years in the future and the models that you mentioned doesn't work very well in cases where it has to do a multistep forecast for so many steps",
      "votes": null
    },
    {
      "id": "677914",
      "postDate": "11/20/2019 18:55:58",
      "content": "<p>Thanks!\nIs there any other approach I can try using other than LGBM and considering it as a Standard Regression problem? (ARIMA maybe?)</p>",
      "rawMarkdown": "Thanks!\nIs there any other approach I can try using other than LGBM and considering it as a Standard Regression problem? (ARIMA maybe?)",
      "votes": null
    },
    {
      "id": "682614",
      "postDate": "11/27/2019 17:16:58",
      "content": "<p>Honestly speaking, its not really a time series competition. If you submit the train data meter readings as your test data predictions, you get a pretty good score! This means that the test data is very similar to train data and doesn't really change over the years. </p>\n\n<p>If you see some EDA kernels that have test data plotted, the plots for the train and test look identical. Therefore, a gradient boosting model would be able to fully exploit this similarity as compared to a time series model. </p>\n\n<p>Moreover, there is only 1 year of training data, so time series models won't be able to learn the 'trend' over the years.\nARMIA is a relatively outdated algorithm, and won't be able to learn the 'trend' which I mentioned above. For example, if only 1-year data is used then the following may happen (it is just an example):\n1. Meter readings are high in November and December due to thanksgiving and Christmas.\n2. ARIMA learns that meter readings increase as the year progresses.\n3. For test data, ARIMA predicts extremely high values by the time it reaches December of 2018.</p>\n\n<p>So I think its best to stick to gradient boosting models to get the best score possible. (And data leaks, although I haven't used any).</p>",
      "rawMarkdown": "Honestly speaking, its not really a time series competition. If you submit the train data meter readings as your test data predictions, you get a pretty good score! This means that the test data is very similar to train data and doesn't really change over the years. \n\nIf you see some EDA kernels that have test data plotted, the plots for the train and test look identical. Therefore, a gradient boosting model would be able to fully exploit this similarity as compared to a time series model. \n\nMoreover, there is only 1 year of training data, so time series models won't be able to learn the 'trend' over the years.\nARMIA is a relatively outdated algorithm, and won't be able to learn the 'trend' which I mentioned above. For example, if only 1-year data is used then the following may happen (it is just an example):\n1. Meter readings are high in November and December due to thanksgiving and Christmas.\n2. ARIMA learns that meter readings increase as the year progresses.\n3. For test data, ARIMA predicts extremely high values by the time it reaches December of 2018.\n\nSo I think its best to stick to gradient boosting models to get the best score possible. (And data leaks, although I haven't used any).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 677384,
      "author_name": "avikrams",
      "author_url": "",
      "post_date": "11/20/2019 04:47:55",
      "content": "<p>Hi ,  In this competition we have to forecast for 2 years in the future and the models that you mentioned doesn't work very well in cases where it has to do a multistep forecast for so many steps </p>",
      "votes": null,
      "replies": [
        {
          "id": 677914,
          "author_name": "sanjanvijayakumar",
          "author_url": "",
          "post_date": "11/20/2019 18:55:58",
          "content": "<p>Thanks!\nIs there any other approach I can try using other than LGBM and considering it as a Standard Regression problem? (ARIMA maybe?)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 682614,
      "author_name": "yushg123",
      "author_url": "",
      "post_date": "11/27/2019 17:16:58",
      "content": "<p>Honestly speaking, its not really a time series competition. If you submit the train data meter readings as your test data predictions, you get a pretty good score! This means that the test data is very similar to train data and doesn't really change over the years. </p>\n\n<p>If you see some EDA kernels that have test data plotted, the plots for the train and test look identical. Therefore, a gradient boosting model would be able to fully exploit this similarity as compared to a time series model. </p>\n\n<p>Moreover, there is only 1 year of training data, so time series models won't be able to learn the 'trend' over the years.\nARMIA is a relatively outdated algorithm, and won't be able to learn the 'trend' which I mentioned above. For example, if only 1-year data is used then the following may happen (it is just an example):\n1. Meter readings are high in November and December due to thanksgiving and Christmas.\n2. ARIMA learns that meter readings increase as the year progresses.\n3. For test data, ARIMA predicts extremely high values by the time it reaches December of 2018.</p>\n\n<p>So I think its best to stick to gradient boosting models to get the best score possible. (And data leaks, although I haven't used any).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "677276": "Hello!\n\nI noticed a number of notebooks using LightGBM to perform modeling.\nIs there any particular reason for this?\nI did a little bit of research and found that VAR and VARMAX can be used to  model time-series forecasting but did not find any notebooks related to these. \nAm I missing anything here?\nAny help would be appreciated. Thanks!",
    "677384": "Hi ,  In this competition we have to forecast for 2 years in the future and the models that you mentioned doesn't work very well in cases where it has to do a multistep forecast for so many steps",
    "677914": "Thanks!\nIs there any other approach I can try using other than LGBM and considering it as a Standard Regression problem? (ARIMA maybe?)",
    "682614": "Honestly speaking, its not really a time series competition. If you submit the train data meter readings as your test data predictions, you get a pretty good score! This means that the test data is very similar to train data and doesn't really change over the years. \n\nIf you see some EDA kernels that have test data plotted, the plots for the train and test look identical. Therefore, a gradient boosting model would be able to fully exploit this similarity as compared to a time series model. \n\nMoreover, there is only 1 year of training data, so time series models won't be able to learn the 'trend' over the years.\nARMIA is a relatively outdated algorithm, and won't be able to learn the 'trend' which I mentioned above. For example, if only 1-year data is used then the following may happen (it is just an example):\n1. Meter readings are high in November and December due to thanksgiving and Christmas.\n2. ARIMA learns that meter readings increase as the year progresses.\n3. For test data, ARIMA predicts extremely high values by the time it reaches December of 2018.\n\nSo I think its best to stick to gradient boosting models to get the best score possible. (And data leaks, although I haven't used any)."
  },
  "source": "meta"
}