The test.csv file contains only the timestamps in August 16, 2017- August 31, 2017. So, when training on the data like transactions or holidays, which are time dependent - only the data within this time boundary would affect the test set performance.
\n\nAnalysis like rise in transactions and sales during Christmas/ New Year or festivals during the rest of the year which would be valuable for improving generalization performance of the model all year round would not be beneficial to this particular time range. So, for time dependent holiday/ transactions data - focus on August. Or am I missing something here?
", "messages": [ { "id": "246499", "postDate": "11/21/2017 07:31:18", "content": "The test.csv file contains only the timestamps in August 16, 2017- August 31, 2017. So, when training on the data like transactions or holidays, which are time dependent - only the data within this time boundary would affect the test set performance.
\n\nAnalysis like rise in transactions and sales during Christmas/ New Year or festivals during the rest of the year which would be valuable for improving generalization performance of the model all year round would not be beneficial to this particular time range. So, for time dependent holiday/ transactions data - focus on August. Or am I missing something here?
", "rawMarkdown": "The test.csv file contains only the timestamps in August 16, 2017- August 31, 2017. So, when training on the data like transactions or holidays, which are time dependent - only the data within this time boundary would affect the test set performance. \n\nAnalysis like rise in transactions and sales during Christmas/ New Year or festivals during the rest of the year which would be valuable for improving generalization performance of the model all year round would not be beneficial to this particular time range. So, for time dependent holiday/ transactions data - focus on August. Or am I missing something here?", "votes": null }, { "id": "247031", "postDate": "11/22/2017 06:26:17", "content": "@inversion I also had this query. Please look into it.
", "rawMarkdown": "inversion I also had this query. Please look into it.", "votes": null }, { "id": "247378", "postDate": "11/22/2017 23:00:20", "content": "So true. If you look at Favorita Analysis Kernel it is based on the second half of August. It would have been more rational to have a test set based on the sales of one year, not sure if the competition designer is regretting not doing this by now, although this would, probably, make forecasting harder.
", "rawMarkdown": "So true. If you look at Favorita Analysis Kernel it is based on the second half of August. It would have been more rational to have a test set based on the sales of one year, not sure if the competition designer is regretting not doing this by now, although this would, probably, make forecasting harder.", "votes": null }, { "id": "247414", "postDate": "11/23/2017 01:58:06", "content": "I asked this in a general thread and the organizers said they only need 2 weeks worth of forward predictions... I still think it's a bit odd, but it's what they want and makes the problem a bit easier (i.e. scripts are performing well with only 2017 training data...)
", "rawMarkdown": "I asked this in a general thread and the organizers said they only need 2 weeks worth of forward predictions... I still think it's a bit odd, but it's what they want and makes the problem a bit easier (i.e. scripts are performing well with only 2017 training data...)", "votes": null } ], "comments": [ { "id": 247031, "author_name": "bhagyeshvikani", "author_url": "", "post_date": "11/22/2017 06:26:17", "content": "@inversion I also had this query. Please look into it.
", "votes": null, "replies": [] }, { "id": 247378, "author_name": "rawi707", "author_url": "", "post_date": "11/22/2017 23:00:20", "content": "So true. If you look at Favorita Analysis Kernel it is based on the second half of August. It would have been more rational to have a test set based on the sales of one year, not sure if the competition designer is regretting not doing this by now, although this would, probably, make forecasting harder.
", "votes": null, "replies": [] }, { "id": 247414, "author_name": "happycube", "author_url": "", "post_date": "11/23/2017 01:58:06", "content": "I asked this in a general thread and the organizers said they only need 2 weeks worth of forward predictions... I still think it's a bit odd, but it's what they want and makes the problem a bit easier (i.e. scripts are performing well with only 2017 training data...)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "246499": "The test.csv file contains only the timestamps in August 16, 2017- August 31, 2017. So, when training on the data like transactions or holidays, which are time dependent - only the data within this time boundary would affect the test set performance. \n\nAnalysis like rise in transactions and sales during Christmas/ New Year or festivals during the rest of the year which would be valuable for improving generalization performance of the model all year round would not be beneficial to this particular time range. So, for time dependent holiday/ transactions data - focus on August. Or am I missing something here?", "247031": "inversion I also had this query. Please look into it.", "247378": "So true. If you look at Favorita Analysis Kernel it is based on the second half of August. It would have been more rational to have a test set based on the sales of one year, not sure if the competition designer is regretting not doing this by now, although this would, probably, make forecasting harder.", "247414": "I asked this in a general thread and the organizers said they only need 2 weeks worth of forward predictions... I still think it's a bit odd, but it's what they want and makes the problem a bit easier (i.e. scripts are performing well with only 2017 training data...)" }, "source": "meta" }