I always have this question when building a time series model with multiple segment. Should I build one time series model for all meter types or build multiple models per meter types and aggregate the result? Below is my understanding for the pro and con for each method.
\n\nMethod 1: Build one time series model for all meter types, overall prediction may be more accurate because error are offset among segments, but RMSE may be bigger among segment\n Method 2: Build multiple models per meter types and aggregate the result, RMSE among each segments may be low but overall RMSE may be off.
\n\nSo what are the trade-off among each method, and what are some of the remediation of each method. Are there better way to resolve the issues? I hope that the Kaggle community can help me to understand this problem more.
", "messages": [ { "id": "659279", "postDate": "10/27/2019 11:56:22", "content": "I always have this question when building a time series model with multiple segment. Should I build one time series model for all meter types or build multiple models per meter types and aggregate the result? Below is my understanding for the pro and con for each method.
\n\nMethod 1: Build one time series model for all meter types, overall prediction may be more accurate because error are offset among segments, but RMSE may be bigger among segment\n Method 2: Build multiple models per meter types and aggregate the result, RMSE among each segments may be low but overall RMSE may be off.
\n\nSo what are the trade-off among each method, and what are some of the remediation of each method. Are there better way to resolve the issues? I hope that the Kaggle community can help me to understand this problem more.
", "rawMarkdown": "I always have this question when building a time series model with multiple segment. Should I build one time series model for all meter types or build multiple models per meter types and aggregate the result? Below is my understanding for the pro and con for each method.\n\nMethod 1: Build one time series model for all meter types, overall prediction may be more accurate because error are offset among segments, but RMSE may be bigger among segment\n Method 2: Build multiple models per meter types and aggregate the result, RMSE among each segments may be low but overall RMSE may be off.\n\nSo what are the trade-off among each method, and what are some of the remediation of each method. Are there better way to resolve the issues? I hope that the Kaggle community can help me to understand this problem more.", "votes": null }, { "id": "659293", "postDate": "10/27/2019 12:19:12", "content": "There's never a satisfying answer to these sort of questions. There is no \"best way\" to do things. Competitive data science is very a field where you have to try things out for yourself. As Nike says, just do it. That being said, tree models will recognize that splitting by meter helps a lot and will take care of this for you, so building separate models probably won't help by a significant margin.
", "rawMarkdown": "There's never a satisfying answer to these sort of questions. There is no \"best way\" to do things. Competitive data science is very a field where you have to try things out for yourself. As Nike says, *just do it*. That being said, tree models will recognize that splitting by meter helps a lot and will take care of this for you, so building separate models probably won't help by a significant margin.", "votes": null }, { "id": "659379", "postDate": "10/27/2019 14:53:14", "content": "Totally agree with you.
", "rawMarkdown": "Totally agree with you.", "votes": null } ], "comments": [ { "id": 659293, "author_name": "maxhalford", "author_url": "", "post_date": "10/27/2019 12:19:12", "content": "There's never a satisfying answer to these sort of questions. There is no \"best way\" to do things. Competitive data science is very a field where you have to try things out for yourself. As Nike says, just do it. That being said, tree models will recognize that splitting by meter helps a lot and will take care of this for you, so building separate models probably won't help by a significant margin.
", "votes": null, "replies": [ { "id": 659379, "author_name": "klloke", "author_url": "", "post_date": "10/27/2019 14:53:14", "content": "Totally agree with you.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "659279": "I always have this question when building a time series model with multiple segment. Should I build one time series model for all meter types or build multiple models per meter types and aggregate the result? Below is my understanding for the pro and con for each method.\n\nMethod 1: Build one time series model for all meter types, overall prediction may be more accurate because error are offset among segments, but RMSE may be bigger among segment\n Method 2: Build multiple models per meter types and aggregate the result, RMSE among each segments may be low but overall RMSE may be off.\n\nSo what are the trade-off among each method, and what are some of the remediation of each method. Are there better way to resolve the issues? I hope that the Kaggle community can help me to understand this problem more.", "659293": "There's never a satisfying answer to these sort of questions. There is no \"best way\" to do things. Competitive data science is very a field where you have to try things out for yourself. As Nike says, *just do it*. That being said, tree models will recognize that splitting by meter helps a lot and will take care of this for you, so building separate models probably won't help by a significant margin.", "659379": "Totally agree with you." }, "source": "meta" }