Hard to say now whether approaches that flatten the data or treat it via time series methods will be dominant (I'd bet the real answer is an ensemble of both!), but neural nets provide a natural class of model for handling a large set of related time series simultaneously without needing to select series-specific parameters. I've put together some walkthroughs and explanations of this style of modeling (using Keras) that you can find here on github.
\n\nMy work is focused on WaveNet, which I believe is one of the most efficient forecasting architectures out there, but also includes a notebook on LSTM encoder-decoder modeling. My hope is that the notebooks are a good technical introduction to this topic and might be a reasonable jumping off point for building a neural net to apply to this competition.
\n\nIf you want a higher level conceptual overview, some of my blog posts may be of interest too.
\n\nHope this is helpful and good luck with the competition!
", "messages": [ { "id": 650597, "postDate": "2019-10-16T14:31:55.627Z", "content": "Hard to say now whether approaches that flatten the data or treat it via time series methods will be dominant (I'd bet the real answer is an ensemble of both!), but neural nets provide a natural class of model for handling a large set of related time series simultaneously without needing to select series-specific parameters. I've put together some walkthroughs and explanations of this style of modeling (using Keras) that you can find here on github.
\n\nMy work is focused on WaveNet, which I believe is one of the most efficient forecasting architectures out there, but also includes a notebook on LSTM encoder-decoder modeling. My hope is that the notebooks are a good technical introduction to this topic and might be a reasonable jumping off point for building a neural net to apply to this competition.
\n\nIf you want a higher level conceptual overview, some of my blog posts may be of interest too.
\n\nHope this is helpful and good luck with the competition!
", "rawMarkdown": "Hard to say now whether approaches that flatten the data or treat it via time series methods will be dominant (I'd bet the real answer is an ensemble of both!), but neural nets provide a natural class of model for handling a large set of related time series simultaneously without needing to select series-specific parameters. I've put together some walkthroughs and explanations of this style of modeling (using Keras) that you can find [here on github](https://github.com/JEddy92/TimeSeries_Seq2Seq). \n\nMy work is focused on WaveNet, which I believe is one of the most efficient forecasting architectures out there, but also includes a notebook on LSTM encoder-decoder modeling. My hope is that the notebooks are a good technical introduction to this topic and might be a reasonable jumping off point for building a neural net to apply to this competition.\n\nIf you want a higher level conceptual overview, some of my [blog posts](https://jeddy92.github.io/) may be of interest too.\n\nHope this is helpful and good luck with the competition!", "votes": 63 }, { "id": 650802, "postDate": "2019-10-16T17:37:23.247Z", "content": "I have implemented your temporal convolutional networks at my job in forecasting electricity usage, and it outperforms lgbm. I highly recommend people read his blogs & github. Thanks jeddy
", "rawMarkdown": "I have implemented your temporal convolutional networks at my job in forecasting electricity usage, and it outperforms lgbm. I highly recommend people read his blogs & github. Thanks jeddy", "votes": 3 }, { "id": 655276, "postDate": "2019-10-22T21:36:24.347Z", "content": "thanks for sharing.\nPS: there are no bad architectures, could be worse data science knowledge. That's why we are here.
", "rawMarkdown": "thanks for sharing.\nPS: there are no bad architectures, could be worse data science knowledge. That's why we are here. ", "votes": 1 }, { "id": 652558, "postDate": "2019-10-19T02:57:16.587Z", "content": "I read your github too, that is a great tutorial!
", "rawMarkdown": "I read your github too, that is a great tutorial!" }, { "id": 677785, "postDate": "2019-11-20T15:32:39.220Z", "content": "So if I understand correctly this method does not use any explanatory variables other than the past values of the variable we want to predict? Do you think this could work in our case where the prediction horizon is longer than a year?
", "rawMarkdown": "So if I understand correctly this method does not use any explanatory variables other than the past values of the variable we want to predict? Do you think this could work in our case where the prediction horizon is longer than a year?" }, { "id": 661354, "postDate": "2019-10-30T07:19:41.413Z", "content": "Hi, your github toturial is great, i'm curious about wavenet performance in this competition, would you share wavenet PB score？
", "rawMarkdown": "Hi, your github toturial is great, i'm curious about wavenet performance in this competition, would you share wavenet PB score？" }, { "id": 667158, "postDate": "2019-11-06T21:59:28.357Z", "rawMarkdown": "", "isDeleted": true }, { "id": 655181, "postDate": "2019-10-22T19:07:13.350Z", "content": "Thanks a lot Joe, as every time !
", "rawMarkdown": "Thanks a lot Joe, as every time !", "votes": 1 }, { "id": 650693, "postDate": "2019-10-16T15:49:14.460Z", "content": "Great.. Very Helpful\nThanks @aquatic
", "rawMarkdown": "Great.. Very Helpful\nThanks @aquatic ", "votes": 1 }, { "id": 654103, "postDate": "2019-10-21T12:52:32.330Z", "content": "Nice! Thanks for sharing!
", "rawMarkdown": "Nice! Thanks for sharing!", "votes": 2 }, { "id": 655362, "postDate": "2019-10-23T01:21:13.057Z", "content": "Thanks for sharing :)
", "rawMarkdown": "Thanks for sharing :)" } ], "comments": [ { "id": 650802, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2019-10-16T17:37:23.247000", "content": "I have implemented your temporal convolutional networks at my job in forecasting electricity usage, and it outperforms lgbm. I highly recommend people read his blogs & github. Thanks jeddy
", "votes": 3, "replies": [] }, { "id": 655276, "author_name": "Adrian Zinovei", "author_url": "", "post_date": "2019-10-22T21:36:24.347000", "content": "thanks for sharing.\nPS: there are no bad architectures, could be worse data science knowledge. That's why we are here.
", "votes": 1, "replies": [] }, { "id": 652558, "author_name": "LongYin/杰少", "author_url": "", "post_date": "2019-10-19T02:57:16.587000", "content": "I read your github too, that is a great tutorial!
", "votes": 0, "replies": [] }, { "id": 677785, "author_name": "Benedetto Grillone", "author_url": "", "post_date": "2019-11-20T15:32:39.220000", "content": "So if I understand correctly this method does not use any explanatory variables other than the past values of the variable we want to predict? Do you think this could work in our case where the prediction horizon is longer than a year?
", "votes": 0, "replies": [] }, { "id": 661354, "author_name": "KKY", "author_url": "", "post_date": "2019-10-30T07:19:41.413000", "content": "Hi, your github toturial is great, i'm curious about wavenet performance in this competition, would you share wavenet PB score？
", "votes": 0, "replies": [] }, { "id": 667158, "author_name": "", "author_url": "", "post_date": "2019-11-06T21:59:28.357000", "content": "", "votes": 0, "replies": [] }, { "id": 655181, "author_name": "mezoganet", "author_url": "", "post_date": "2019-10-22T19:07:13.350000", "content": "Thanks a lot Joe, as every time !
", "votes": 1, "replies": [] }, { "id": 650693, "author_name": "Ailurophile", "author_url": "", "post_date": "2019-10-16T15:49:14.460000", "content": "Great.. Very Helpful\nThanks @aquatic
", "votes": 1, "replies": [] }, { "id": 654103, "author_name": "Carlo", "author_url": "", "post_date": "2019-10-21T12:52:32.330000", "content": "Nice! Thanks for sharing!
", "votes": 2, "replies": [] }, { "id": 655362, "author_name": "ANAPARK", "author_url": "", "post_date": "2019-10-23T01:21:13.057000", "content": "Thanks for sharing :)
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "650597": "Hard to say now whether approaches that flatten the data or treat it via time series methods will be dominant (I'd bet the real answer is an ensemble of both!), but neural nets provide a natural class of model for handling a large set of related time series simultaneously without needing to select series-specific parameters. I've put together some walkthroughs and explanations of this style of modeling (using Keras) that you can find [here on github](https://github.com/JEddy92/TimeSeries_Seq2Seq). \n\nMy work is focused on WaveNet, which I believe is one of the most efficient forecasting architectures out there, but also includes a notebook on LSTM encoder-decoder modeling. My hope is that the notebooks are a good technical introduction to this topic and might be a reasonable jumping off point for building a neural net to apply to this competition.\n\nIf you want a higher level conceptual overview, some of my [blog posts](https://jeddy92.github.io/) may be of interest too.\n\nHope this is helpful and good luck with the competition!", "650802": "I have implemented your temporal convolutional networks at my job in forecasting electricity usage, and it outperforms lgbm. I highly recommend people read his blogs & github. Thanks jeddy", "655276": "thanks for sharing.\nPS: there are no bad architectures, could be worse data science knowledge. That's why we are here. ", "652558": "I read your github too, that is a great tutorial!", "677785": "So if I understand correctly this method does not use any explanatory variables other than the past values of the variable we want to predict? Do you think this could work in our case where the prediction horizon is longer than a year?", "661354": "Hi, your github toturial is great, i'm curious about wavenet performance in this competition, would you share wavenet PB score？", "667158": "", "655181": "Thanks a lot Joe, as every time !", "650693": "Great.. Very Helpful\nThanks @aquatic ", "654103": "Nice! Thanks for sharing!", "655362": "Thanks for sharing :)" } }