{
  "id": 71547,
  "title": "Is this a Timeseries Forecasting Problem ?",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/71547",
  "author_name": "",
  "post_date": "2018-11-14T17:11:18.499148Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hey community, can we think problem as Timeseries Forecasting problem where forecasting is done on <strong>fullVisitorId</strong> level ?</p>\n\n<p>Thanks in advance.</p>",
  "messages": [
    {
      "id": "421178",
      "postDate": "11/14/2018 17:11:18",
      "content": "<p>Hey community, can we think problem as Timeseries Forecasting problem where forecasting is done on <strong>fullVisitorId</strong> level ?</p>\n\n<p>Thanks in advance.</p>",
      "rawMarkdown": "Hey community, can we think problem as Timeseries Forecasting problem where forecasting is done on **fullVisitorId** level ?\n\nThanks in advance.",
      "votes": null
    },
    {
      "id": "421197",
      "postDate": "11/14/2018 17:44:56",
      "content": "<p>Hi Praneeth,</p>\n\n<p>I think a TimeSeries has a few conditions, which the dataset doesn't meet.\nWikipedia:\nA time series is a series of data points indexed (or listed or graphed) in time order. Most commonly, a time series is a sequence taken at successive equally spaced points in time. Thus it is a sequence of discrete-time data</p>\n\n<p>The timevalues aren't unique, nor are they a sequence of equally spaced points in time. So it would take much editing to make this a Timeseries forecast problem and probably would miss the point of the competition if set up that way.</p>",
      "rawMarkdown": "Hi Praneeth,\n\nI think a TimeSeries has a few conditions, which the dataset doesn't meet.\nWikipedia:\nA time series is a series of data points indexed (or listed or graphed) in time order. Most commonly, a time series is a sequence taken at successive equally spaced points in time. Thus it is a sequence of discrete-time data\n\nThe timevalues aren't unique, nor are they a sequence of equally spaced points in time. So it would take much editing to make this a Timeseries forecast problem and probably would miss the point of the competition if set up that way.",
      "votes": null
    },
    {
      "id": "430403",
      "postDate": "11/30/2018 10:32:22",
      "content": "<p>I believe it is better to see this problem as a time series forecasting problem since, for the time period the private leaderboard will not depend on all the features that are till october. Since you dont have those features for december and jan whatever the public kernels are predicting are very misleading.\nWhat predictive modelling approach will mean that you are predicting for fullVisitorId for the month of september and october.\nSo its evaluation in private leader board will mean that whatever you predicted for these two months how close will they be for the month of december and january. And if you explore the data at a fullVisitorId level, you will see the chances of that happening are very low, which is why many people are betting on all zero predictions.</p>\n\n<p>Time series perspective is better for this problem, but if you explore data at fullVisitorId level you will see that time series modelling wont give you much of results as the data is highly intermittent and volatile, so it has just become a lottery or betting game with two main questions:\n1. Which fullVisitorId is having better chances of doing transactions in the month of december and january</p>\n\n<ol>\n<li>All the fullVisitorId which are having good chances of transactions, how much transactions will they make in those two months</li>\n</ol>",
      "rawMarkdown": "I believe it is better to see this problem as a time series forecasting problem since, for the time period the private leaderboard will not depend on all the features that are till october. Since you dont have those features for december and jan whatever the public kernels are predicting are very misleading.\nWhat predictive modelling approach will mean that you are predicting for fullVisitorId for the month of september and october.\nSo its evaluation in private leader board will mean that whatever you predicted for these two months how close will they be for the month of december and january. And if you explore the data at a fullVisitorId level, you will see the chances of that happening are very low, which is why many people are betting on all zero predictions.\n\nTime series perspective is better for this problem, but if you explore data at fullVisitorId level you will see that time series modelling wont give you much of results as the data is highly intermittent and volatile, so it has just become a lottery or betting game with two main questions:\n1. Which fullVisitorId is having better chances of doing transactions in the month of december and january\n\n2. All the fullVisitorId which are having good chances of transactions, how much transactions will they make in those two months",
      "votes": null
    },
    {
      "id": "430411",
      "postDate": "11/30/2018 10:57:19",
      "content": "<p>You can always create time series for every fullVisitorId in test_datav2 not a problem. The point of the competition is already missed. It would have only made sense if they would have rescored kernels like in two sigma stock prediction</p>",
      "rawMarkdown": "You can always create time series for every fullVisitorId in test_datav2 not a problem. The point of the competition is already missed. It would have only made sense if they would have rescored kernels like in two sigma stock prediction",
      "votes": null
    },
    {
      "id": "459984",
      "postDate": "01/22/2019 19:00:55",
      "content": "<p>Yes it is. </p>",
      "rawMarkdown": "Yes it is.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 421197,
      "author_name": "ndekker",
      "author_url": "",
      "post_date": "11/14/2018 17:44:56",
      "content": "<p>Hi Praneeth,</p>\n\n<p>I think a TimeSeries has a few conditions, which the dataset doesn't meet.\nWikipedia:\nA time series is a series of data points indexed (or listed or graphed) in time order. Most commonly, a time series is a sequence taken at successive equally spaced points in time. Thus it is a sequence of discrete-time data</p>\n\n<p>The timevalues aren't unique, nor are they a sequence of equally spaced points in time. So it would take much editing to make this a Timeseries forecast problem and probably would miss the point of the competition if set up that way.</p>",
      "votes": null,
      "replies": [
        {
          "id": 430411,
          "author_name": "shubhinderrana",
          "author_url": "",
          "post_date": "11/30/2018 10:57:19",
          "content": "<p>You can always create time series for every fullVisitorId in test_datav2 not a problem. The point of the competition is already missed. It would have only made sense if they would have rescored kernels like in two sigma stock prediction</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 430403,
      "author_name": "shubhinderrana",
      "author_url": "",
      "post_date": "11/30/2018 10:32:22",
      "content": "<p>I believe it is better to see this problem as a time series forecasting problem since, for the time period the private leaderboard will not depend on all the features that are till october. Since you dont have those features for december and jan whatever the public kernels are predicting are very misleading.\nWhat predictive modelling approach will mean that you are predicting for fullVisitorId for the month of september and october.\nSo its evaluation in private leader board will mean that whatever you predicted for these two months how close will they be for the month of december and january. And if you explore the data at a fullVisitorId level, you will see the chances of that happening are very low, which is why many people are betting on all zero predictions.</p>\n\n<p>Time series perspective is better for this problem, but if you explore data at fullVisitorId level you will see that time series modelling wont give you much of results as the data is highly intermittent and volatile, so it has just become a lottery or betting game with two main questions:\n1. Which fullVisitorId is having better chances of doing transactions in the month of december and january</p>\n\n<ol>\n<li>All the fullVisitorId which are having good chances of transactions, how much transactions will they make in those two months</li>\n</ol>",
      "votes": null,
      "replies": []
    },
    {
      "id": 459984,
      "author_name": "paradoxlover",
      "author_url": "",
      "post_date": "01/22/2019 19:00:55",
      "content": "<p>Yes it is. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "421178": "Hey community, can we think problem as Timeseries Forecasting problem where forecasting is done on **fullVisitorId** level ?\n\nThanks in advance.",
    "421197": "Hi Praneeth,\n\nI think a TimeSeries has a few conditions, which the dataset doesn't meet.\nWikipedia:\nA time series is a series of data points indexed (or listed or graphed) in time order. Most commonly, a time series is a sequence taken at successive equally spaced points in time. Thus it is a sequence of discrete-time data\n\nThe timevalues aren't unique, nor are they a sequence of equally spaced points in time. So it would take much editing to make this a Timeseries forecast problem and probably would miss the point of the competition if set up that way.",
    "430403": "I believe it is better to see this problem as a time series forecasting problem since, for the time period the private leaderboard will not depend on all the features that are till october. Since you dont have those features for december and jan whatever the public kernels are predicting are very misleading.\nWhat predictive modelling approach will mean that you are predicting for fullVisitorId for the month of september and october.\nSo its evaluation in private leader board will mean that whatever you predicted for these two months how close will they be for the month of december and january. And if you explore the data at a fullVisitorId level, you will see the chances of that happening are very low, which is why many people are betting on all zero predictions.\n\nTime series perspective is better for this problem, but if you explore data at fullVisitorId level you will see that time series modelling wont give you much of results as the data is highly intermittent and volatile, so it has just become a lottery or betting game with two main questions:\n1. Which fullVisitorId is having better chances of doing transactions in the month of december and january\n\n2. All the fullVisitorId which are having good chances of transactions, how much transactions will they make in those two months",
    "430411": "You can always create time series for every fullVisitorId in test_datav2 not a problem. The point of the competition is already missed. It would have only made sense if they would have rescored kernels like in two sigma stock prediction",
    "459984": "Yes it is."
  },
  "source": "meta"
}