{
  "id": 81531,
  "title": "Score of Sample Submission",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/81531",
  "author_name": "",
  "post_date": "2019-02-22T07:39:51.981622200Z",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>According to my late submission, sample submission (all 0 prediction) is scored 0.88843 in private LB. \nHowever in my local evaluation, all 0 predictions were scored around 0.15 ~ 0.35. \n<a href=\"https://www.kaggle.com/mtfall/all-0-prediction-score-check\">https://www.kaggle.com/mtfall/all-0-prediction-score-check</a></p>\n\n<p>Is there any mistake in my code? or, did GStore great business on period of December 1st 2018 through January 31st 2019?</p>",
  "messages": [
    {
      "id": "476470",
      "postDate": "02/22/2019 07:39:51",
      "content": "<p>According to my late submission, sample submission (all 0 prediction) is scored 0.88843 in private LB. \nHowever in my local evaluation, all 0 predictions were scored around 0.15 ~ 0.35. \n<a href=\"https://www.kaggle.com/mtfall/all-0-prediction-score-check\">https://www.kaggle.com/mtfall/all-0-prediction-score-check</a></p>\n\n<p>Is there any mistake in my code? or, did GStore great business on period of December 1st 2018 through January 31st 2019?</p>",
      "rawMarkdown": "According to my late submission, sample submission (all 0 prediction) is scored 0.88843 in private LB. \nHowever in my local evaluation, all 0 predictions were scored around 0.15 ~ 0.35. \nhttps://www.kaggle.com/mtfall/all-0-prediction-score-check\n\nIs there any mistake in my code? or, did GStore great business on period of December 1st 2018 through January 31st 2019?",
      "votes": null
    },
    {
      "id": "476634",
      "postDate": "02/22/2019 13:02:11",
      "content": "<p>More likely there was some non random sampling in train dataset. (As comented in other threads, it also explains weird lack of seasonality)</p>",
      "rawMarkdown": "More likely there was some non random sampling in train dataset. (As comented in other threads, it also explains weird lack of seasonality)",
      "votes": null
    },
    {
      "id": "476710",
      "postDate": "02/22/2019 15:24:45",
      "content": "<p>Thanks for comment and congrats solo gold, miguel. \nI read your solution but I can't understand why you think there were sampling process in training data. Can you explain more detail about so called weird lack of seasonality?</p>",
      "rawMarkdown": "Thanks for comment and congrats solo gold, miguel. \nI read your solution but I can't understand why you think there were sampling process in training data. Can you explain more detail about so called weird lack of seasonality?",
      "votes": null
    },
    {
      "id": "476804",
      "postDate": "02/23/2019 07:36:45",
      "content": "<p>Thanks @TakuyaAkiyama, sure I can explain my point; If you aggregate sales data by time you get a time series of sales. This time series exhibits some weekly seasonality, as is normal in most sales data, but no clear yearly seasonality, or trend. And that, while possible, is quite unusual. Sales data very frequently show a big difference in behaviour througout the year because of, well, Christmas, Summertime...  So there is the weirdness, train data is quite flat.</p>\n\n<p>If train data was not sampled and test period was supposed to have the same nature of train... then next suspect for that big difference in scores would be test period being different to train period, what would be an anomaly. But I bet more on the non random sampling of train data.</p>\n\n<p>Oh, just one last explanation about the -yearly- seasonality thing. I commented in the solution that if there had been seasonality it was possibly game over. That is not competely true, as one can try to remove the seasonal component and add it on prediction. But two years of data is not enough to reliably do that, hence the comment on \"non yearly seasonality == good \". It allowed to reuse time windows without worries.</p>\n\n<p>Hope that was more clear :-) </p>",
      "rawMarkdown": "Thanks @TakuyaAkiyama, sure I can explain my point; If you aggregate sales data by time you get a time series of sales. This time series exhibits some weekly seasonality, as is normal in most sales data, but no clear yearly seasonality, or trend. And that, while possible, is quite unusual. Sales data very frequently show a big difference in behaviour througout the year because of, well, Christmas, Summertime...  So there is the weirdness, train data is quite flat.\n\nIf train data was not sampled and test period was supposed to have the same nature of train... then next suspect for that big difference in scores would be test period being different to train period, what would be an anomaly. But I bet more on the non random sampling of train data.\n\nOh, just one last explanation about the -yearly- seasonality thing. I commented in the solution that if there had been seasonality it was possibly game over. That is not competely true, as one can try to remove the seasonal component and add it on prediction. But two years of data is not enough to reliably do that, hence the comment on \"non yearly seasonality == good \". It allowed to reuse time windows without worries.\n\nHope that was more clear :-)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 476634,
      "author_name": "miguelpm",
      "author_url": "",
      "post_date": "02/22/2019 13:02:11",
      "content": "<p>More likely there was some non random sampling in train dataset. (As comented in other threads, it also explains weird lack of seasonality)</p>",
      "votes": null,
      "replies": [
        {
          "id": 476710,
          "author_name": "mtfall",
          "author_url": "",
          "post_date": "02/22/2019 15:24:45",
          "content": "<p>Thanks for comment and congrats solo gold, miguel. \nI read your solution but I can't understand why you think there were sampling process in training data. Can you explain more detail about so called weird lack of seasonality?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 476804,
          "author_name": "miguelpm",
          "author_url": "",
          "post_date": "02/23/2019 07:36:45",
          "content": "<p>Thanks @TakuyaAkiyama, sure I can explain my point; If you aggregate sales data by time you get a time series of sales. This time series exhibits some weekly seasonality, as is normal in most sales data, but no clear yearly seasonality, or trend. And that, while possible, is quite unusual. Sales data very frequently show a big difference in behaviour througout the year because of, well, Christmas, Summertime...  So there is the weirdness, train data is quite flat.</p>\n\n<p>If train data was not sampled and test period was supposed to have the same nature of train... then next suspect for that big difference in scores would be test period being different to train period, what would be an anomaly. But I bet more on the non random sampling of train data.</p>\n\n<p>Oh, just one last explanation about the -yearly- seasonality thing. I commented in the solution that if there had been seasonality it was possibly game over. That is not competely true, as one can try to remove the seasonal component and add it on prediction. But two years of data is not enough to reliably do that, hence the comment on \"non yearly seasonality == good \". It allowed to reuse time windows without worries.</p>\n\n<p>Hope that was more clear :-) </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "476470": "According to my late submission, sample submission (all 0 prediction) is scored 0.88843 in private LB. \nHowever in my local evaluation, all 0 predictions were scored around 0.15 ~ 0.35. \nhttps://www.kaggle.com/mtfall/all-0-prediction-score-check\n\nIs there any mistake in my code? or, did GStore great business on period of December 1st 2018 through January 31st 2019?",
    "476634": "More likely there was some non random sampling in train dataset. (As comented in other threads, it also explains weird lack of seasonality)",
    "476710": "Thanks for comment and congrats solo gold, miguel. \nI read your solution but I can't understand why you think there were sampling process in training data. Can you explain more detail about so called weird lack of seasonality?",
    "476804": "Thanks @TakuyaAkiyama, sure I can explain my point; If you aggregate sales data by time you get a time series of sales. This time series exhibits some weekly seasonality, as is normal in most sales data, but no clear yearly seasonality, or trend. And that, while possible, is quite unusual. Sales data very frequently show a big difference in behaviour througout the year because of, well, Christmas, Summertime...  So there is the weirdness, train data is quite flat.\n\nIf train data was not sampled and test period was supposed to have the same nature of train... then next suspect for that big difference in scores would be test period being different to train period, what would be an anomaly. But I bet more on the non random sampling of train data.\n\nOh, just one last explanation about the -yearly- seasonality thing. I commented in the solution that if there had been seasonality it was possibly game over. That is not competely true, as one can try to remove the seasonal component and add it on prediction. But two years of data is not enough to reliably do that, hence the comment on \"non yearly seasonality == good \". It allowed to reuse time windows without worries.\n\nHope that was more clear :-)"
  },
  "source": "meta"
}