For this competition and for predicting time series in general, you will surely find useful this paper:
\nJanuschowski, Tim, et al. \"Forecasting with trees.\" International Journal of Forecasting (2021)
\nhttps://www.sciencedirect.com/science/article/pii/S0169207021001679
It explains about the recent success of gradient boosting methods in time series predictions and why it is here to stay. You will surely find illuminating considerations such as:
\n\"tree-based models can very successfully be used as black-box learners. This allows contestants to focus on improving the model input (through feature engineering) and optimizing their ranking through well-understood recipes, such as careful hyperparameter tuning and ensembling. Feature engineering is arguably both more accessible and more composable than the building blocks of deep learning-based forecasting techniques\"
\nEnjoy reading! You'll find many useful ideas in the paper. Happy learning!
", "messages": [ { "id": "1634153", "postDate": "12/31/2021 11:28:34", "content": "For this competition and for predicting time series in general, you will surely find useful this paper:
\nJanuschowski, Tim, et al. \"Forecasting with trees.\" International Journal of Forecasting (2021)
\nhttps://www.sciencedirect.com/science/article/pii/S0169207021001679
It explains about the recent success of gradient boosting methods in time series predictions and why it is here to stay. You will surely find illuminating considerations such as:
\n\"tree-based models can very successfully be used as black-box learners. This allows contestants to focus on improving the model input (through feature engineering) and optimizing their ranking through well-understood recipes, such as careful hyperparameter tuning and ensembling. Feature engineering is arguably both more accessible and more composable than the building blocks of deep learning-based forecasting techniques\"
\nEnjoy reading! You'll find many useful ideas in the paper. Happy learning!
", "rawMarkdown": "For this competition and for predicting time series in general, you will surely find useful this paper:\n\n**Januschowski, Tim, et al. \"Forecasting with trees.\" International Journal of Forecasting (2021)**\nhttps://www.sciencedirect.com/science/article/pii/S0169207021001679\n\nIt explains about the recent success of gradient boosting methods in time series predictions and why it is here to stay. You will surely find illuminating considerations such as:\n\n\"tree-based models can very successfully be used as black-box learners. This allows contestants to focus on improving the model input (through feature engineering) and optimizing their ranking through well-understood recipes, such as careful hyperparameter tuning and ensembling. Feature engineering is arguably both more accessible and more composable than the building blocks of deep learning-based forecasting techniques\"\n\nEnjoy reading! You'll find many useful ideas in the paper. Happy learning!", "votes": null }, { "id": "1634188", "postDate": "12/31/2021 11:47:29", "content": "Really needed an article for time series! Thanks for sharing! Keep inspiring @lucamassaron
", "rawMarkdown": "Really needed an article for time series! Thanks for sharing! Keep inspiring @lucamassaron", "votes": null }, { "id": "1635217", "postDate": "01/01/2022 15:42:07", "content": "Thanks @lucamassaron for sharing this.
", "rawMarkdown": "Thanks @lucamassaron for sharing this.", "votes": null }, { "id": "1639375", "postDate": "01/05/2022 15:33:51", "content": "Thanks for sharing. I enjoyed reading and agree that there are many useful ideas here both for this competition and for time-series forecasting in general. The paper is based on the M5 Forecasting competition and the authors note some surprise that many of the best-performing models were tree-based rather than Deep Learning based. Based on what they wrote it seems that some approaches are more popular in academia than others but for pragmatic implementations and competitions, those are not always the most effective. One example of this is that they note post-processing can be particularly effective and hard to beat in some cases even though the purists recognize that it is only needed due to shortcomings in the models and prefer not to use it if possible.
", "rawMarkdown": "Thanks for sharing. I enjoyed reading and agree that there are many useful ideas here both for this competition and for time-series forecasting in general. The paper is based on the M5 Forecasting competition and the authors note some surprise that many of the best-performing models were tree-based rather than Deep Learning based. Based on what they wrote it seems that some approaches are more popular in academia than others but for pragmatic implementations and competitions, those are not always the most effective. One example of this is that they note post-processing can be particularly effective and hard to beat in some cases even though the purists recognize that it is only needed due to shortcomings in the models and prefer not to use it if possible.", "votes": null } ], "comments": [ { "id": 1634188, "author_name": "ishaanthareja007", "author_url": "", "post_date": "12/31/2021 11:47:29", "content": "Really needed an article for time series! Thanks for sharing! Keep inspiring @lucamassaron
", "votes": null, "replies": [] }, { "id": 1635217, "author_name": "bandiang2", "author_url": "", "post_date": "01/01/2022 15:42:07", "content": "Thanks @lucamassaron for sharing this.
", "votes": null, "replies": [] }, { "id": 1639375, "author_name": "datahobbit", "author_url": "", "post_date": "01/05/2022 15:33:51", "content": "Thanks for sharing. I enjoyed reading and agree that there are many useful ideas here both for this competition and for time-series forecasting in general. The paper is based on the M5 Forecasting competition and the authors note some surprise that many of the best-performing models were tree-based rather than Deep Learning based. Based on what they wrote it seems that some approaches are more popular in academia than others but for pragmatic implementations and competitions, those are not always the most effective. One example of this is that they note post-processing can be particularly effective and hard to beat in some cases even though the purists recognize that it is only needed due to shortcomings in the models and prefer not to use it if possible.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1634153": "For this competition and for predicting time series in general, you will surely find useful this paper:\n\n**Januschowski, Tim, et al. \"Forecasting with trees.\" International Journal of Forecasting (2021)**\nhttps://www.sciencedirect.com/science/article/pii/S0169207021001679\n\nIt explains about the recent success of gradient boosting methods in time series predictions and why it is here to stay. You will surely find illuminating considerations such as:\n\n\"tree-based models can very successfully be used as black-box learners. This allows contestants to focus on improving the model input (through feature engineering) and optimizing their ranking through well-understood recipes, such as careful hyperparameter tuning and ensembling. Feature engineering is arguably both more accessible and more composable than the building blocks of deep learning-based forecasting techniques\"\n\nEnjoy reading! You'll find many useful ideas in the paper. Happy learning!", "1634188": "Really needed an article for time series! Thanks for sharing! Keep inspiring @lucamassaron", "1635217": "Thanks @lucamassaron for sharing this.", "1639375": "Thanks for sharing. I enjoyed reading and agree that there are many useful ideas here both for this competition and for time-series forecasting in general. The paper is based on the M5 Forecasting competition and the authors note some surprise that many of the best-performing models were tree-based rather than Deep Learning based. Based on what they wrote it seems that some approaches are more popular in academia than others but for pragmatic implementations and competitions, those are not always the most effective. One example of this is that they note post-processing can be particularly effective and hard to beat in some cases even though the purists recognize that it is only needed due to shortcomings in the models and prefer not to use it if possible." }, "source": "meta" }