{
  "id": 44013,
  "title": "ARIMA results not as good as simple MA?",
  "url": "/competitions/favorita-grocery-sales-forecasting/discussion/44013",
  "author_name": "",
  "post_date": "2017-11-22T08:45:03.081663700Z",
  "votes": 4,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi all,\ndid anyone experience that their ARIMA predictions are not as good as a simple MA?</p>\n\n<p>Or am I missing out something...</p>",
  "messages": [
    {
      "id": "247075",
      "postDate": "11/22/2017 08:45:03",
      "content": "<p>Hi all,\ndid anyone experience that their ARIMA predictions are not as good as a simple MA?</p>\n\n<p>Or am I missing out something...</p>",
      "rawMarkdown": "Hi all,\ndid anyone experience that their ARIMA predictions are not as good as a simple MA?\n\nOr am I missing out something...",
      "votes": null
    },
    {
      "id": "247186",
      "postDate": "11/22/2017 14:18:49",
      "content": "<p>me too, so strange</p>",
      "rawMarkdown": "me too, so strange",
      "votes": null
    },
    {
      "id": "247928",
      "postDate": "11/24/2017 10:33:34",
      "content": "<p>Yup, same here.</p>",
      "rawMarkdown": "Yup, same here.",
      "votes": null
    },
    {
      "id": "248030",
      "postDate": "11/24/2017 15:46:39",
      "content": "<p>Time to go Bayesian... or to try using an exogenous variable.</p>",
      "rawMarkdown": "Time to go Bayesian... or to try using an exogenous variable.",
      "votes": null
    },
    {
      "id": "248480",
      "postDate": "11/26/2017 08:34:48",
      "content": "<p>Thank you for suggesting a good idea. I am a novice and I am wondering if you mean ,I can use Bayesian to classify different combinations of store_nbr and item_nbr into different patterns of time series data for ARIMA? Thanks a lot!</p>",
      "rawMarkdown": "Thank you for suggesting a good idea. I am a novice and I am wondering if you mean ,I can use Bayesian to classify different combinations of store_nbr and item_nbr into different patterns of time series data for ARIMA? Thanks a lot!",
      "votes": null
    },
    {
      "id": "249144",
      "postDate": "11/27/2017 20:12:39",
      "content": "<p>Did you guys actually try the auto.arima function ? Because it is supposed to find the optima model, which could be a simpel MA in that case.</p>\n\n<p>I am going to make a submission with that model, shortly - I am quite curious of the result.</p>\n\n<p>From my experience with ARIMA models, however, they are quite bad in predicting horizons bigger than 3-4 days. The reason is that when the ARIMA model sees a peak in sales at the last day, it assumes that for all the remaining days you are going to have an upward trend, and therefore your predictions will be massively overforecasted. For this reason, I suspect that auto.arima will fail miserably with the 14-days horizon of this challenge.</p>",
      "rawMarkdown": "Did you guys actually try the auto.arima function ? Because it is supposed to find the optima model, which could be a simpel MA in that case.\n\nI am going to make a submission with that model, shortly - I am quite curious of the result.\n\nFrom my experience with ARIMA models, however, they are quite bad in predicting horizons bigger than 3-4 days. The reason is that when the ARIMA model sees a peak in sales at the last day, it assumes that for all the remaining days you are going to have an upward trend, and therefore your predictions will be massively overforecasted. For this reason, I suspect that auto.arima will fail miserably with the 14-days horizon of this challenge.",
      "votes": null
    },
    {
      "id": "249257",
      "postDate": "11/28/2017 02:30:53",
      "content": "<p>My experience with using ARIMA in general for univariate time series forecasting is that it underperforms other methods. Auto arima will fit a number of arima models to the time series, trying various combinations of the p (autoregressive) order, d (differencing) order, and q (moving average order) and the use the Bayesian or Akaike information criterion to choose the model with the best fit. Which could certainly turn out to be a simple MA model. </p>\n\n<p>If you want to use arima to handle regular seasonal spikes you will also want to fit the seasonal p,d, and q parameters. Alternatively you could include exogenous regressors like seasonal dummies. </p>\n\n<p>The key is to model spikes whenever possible. Although useful, auto arima is not magical and still requires some econometric knowledge to implement correctly. </p>",
      "rawMarkdown": "My experience with using ARIMA in general for univariate time series forecasting is that it underperforms other methods. Auto arima will fit a number of arima models to the time series, trying various combinations of the p (autoregressive) order, d (differencing) order, and q (moving average order) and the use the Bayesian or Akaike information criterion to choose the model with the best fit. Which could certainly turn out to be a simple MA model. \n\nIf you want to use arima to handle regular seasonal spikes you will also want to fit the seasonal p,d, and q parameters. Alternatively you could include exogenous regressors like seasonal dummies. \n\nThe key is to model spikes whenever possible. Although useful, auto arima is not magical and still requires some econometric knowledge to implement correctly.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 247186,
      "author_name": "chenxinrui",
      "author_url": "",
      "post_date": "11/22/2017 14:18:49",
      "content": "<p>me too, so strange</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 247928,
      "author_name": "konradb",
      "author_url": "",
      "post_date": "11/24/2017 10:33:34",
      "content": "<p>Yup, same here.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 248030,
      "author_name": "puremath86",
      "author_url": "",
      "post_date": "11/24/2017 15:46:39",
      "content": "<p>Time to go Bayesian... or to try using an exogenous variable.</p>",
      "votes": null,
      "replies": [
        {
          "id": 248480,
          "author_name": "chenxinrui",
          "author_url": "",
          "post_date": "11/26/2017 08:34:48",
          "content": "<p>Thank you for suggesting a good idea. I am a novice and I am wondering if you mean ,I can use Bayesian to classify different combinations of store_nbr and item_nbr into different patterns of time series data for ARIMA? Thanks a lot!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 249144,
      "author_name": "ikogias",
      "author_url": "",
      "post_date": "11/27/2017 20:12:39",
      "content": "<p>Did you guys actually try the auto.arima function ? Because it is supposed to find the optima model, which could be a simpel MA in that case.</p>\n\n<p>I am going to make a submission with that model, shortly - I am quite curious of the result.</p>\n\n<p>From my experience with ARIMA models, however, they are quite bad in predicting horizons bigger than 3-4 days. The reason is that when the ARIMA model sees a peak in sales at the last day, it assumes that for all the remaining days you are going to have an upward trend, and therefore your predictions will be massively overforecasted. For this reason, I suspect that auto.arima will fail miserably with the 14-days horizon of this challenge.</p>",
      "votes": null,
      "replies": [
        {
          "id": 249257,
          "author_name": "captcalculator",
          "author_url": "",
          "post_date": "11/28/2017 02:30:53",
          "content": "<p>My experience with using ARIMA in general for univariate time series forecasting is that it underperforms other methods. Auto arima will fit a number of arima models to the time series, trying various combinations of the p (autoregressive) order, d (differencing) order, and q (moving average order) and the use the Bayesian or Akaike information criterion to choose the model with the best fit. Which could certainly turn out to be a simple MA model. </p>\n\n<p>If you want to use arima to handle regular seasonal spikes you will also want to fit the seasonal p,d, and q parameters. Alternatively you could include exogenous regressors like seasonal dummies. </p>\n\n<p>The key is to model spikes whenever possible. Although useful, auto arima is not magical and still requires some econometric knowledge to implement correctly. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "247075": "Hi all,\ndid anyone experience that their ARIMA predictions are not as good as a simple MA?\n\nOr am I missing out something...",
    "247186": "me too, so strange",
    "247928": "Yup, same here.",
    "248030": "Time to go Bayesian... or to try using an exogenous variable.",
    "248480": "Thank you for suggesting a good idea. I am a novice and I am wondering if you mean ,I can use Bayesian to classify different combinations of store_nbr and item_nbr into different patterns of time series data for ARIMA? Thanks a lot!",
    "249144": "Did you guys actually try the auto.arima function ? Because it is supposed to find the optima model, which could be a simpel MA in that case.\n\nI am going to make a submission with that model, shortly - I am quite curious of the result.\n\nFrom my experience with ARIMA models, however, they are quite bad in predicting horizons bigger than 3-4 days. The reason is that when the ARIMA model sees a peak in sales at the last day, it assumes that for all the remaining days you are going to have an upward trend, and therefore your predictions will be massively overforecasted. For this reason, I suspect that auto.arima will fail miserably with the 14-days horizon of this challenge.",
    "249257": "My experience with using ARIMA in general for univariate time series forecasting is that it underperforms other methods. Auto arima will fit a number of arima models to the time series, trying various combinations of the p (autoregressive) order, d (differencing) order, and q (moving average order) and the use the Bayesian or Akaike information criterion to choose the model with the best fit. Which could certainly turn out to be a simple MA model. \n\nIf you want to use arima to handle regular seasonal spikes you will also want to fit the seasonal p,d, and q parameters. Alternatively you could include exogenous regressors like seasonal dummies. \n\nThe key is to model spikes whenever possible. Although useful, auto arima is not magical and still requires some econometric knowledge to implement correctly."
  },
  "source": "meta"
}