Even though there are still two months to go, seeing I'm not a person in finance, I think for me is the moment to do a brief summary on my part and move on, I'm not found of checking lotto every two weeks.
\nFirst of all, I would like to say thank you to my teammate @rafalradwanski for his contribution of generation and selection of features. I cite here his partial work of features of LightGBM model after competition deadline (https://www.kaggle.com/rafalradwanski/g-research-features-to-increase-your-score/notebook), even though we've chosen two notebooks integrated with NBEATS models for final submission, this set of features for LightGBM model deserves a highlight.
\nRegarding two final submissions, I only talk about the simplified version with limited features (features widely shared among public notebooks, not enough with respect to quantity or quality), so I'm more likely to do a resume of what I've learnt after one month's concentrated work. Two main points, the selection of models and the struggle for final submission (especially the last one and a half weeks before deadline).
\nAs I mentioned, I'm not a quant trader, first time participating competition in the domain of finance, so I first read several other Kagglers' share of notebooks, fixed several potential good models, then I started the state of art. Due to the fact that I started quite late, the research period was extreme short. Our team's initial plan was to do practices on two models LightGBM and NBEATS, either the LightGBM or NBEATS has potential to reach a better performance, it depends on topics and how to use it, please refer to (The M5 Accuracy competition: Results, findings and conclusions: https://www.researchgate.net/publication/344487258_The_M5_Accuracy_competition_Results_findings_and_conclusions). Personally, I'm a technical person, NBEATS enables me to practice the technical 'routine' with not so bad result, so I focused on this model. Besides, I was amazed by the design of NBEATS, which motivated me to make more efforts on it (even while I encountered huge difficulties during submissions). Several gains after this practice, the configurations/parameters of training the model, the selection or balance of using data, and the procedure of dealing with the shape (flatten/squeeze or reshape/dimension).
\nAnother aspect I focused most at the late phase of the competition was conquering the submission errors. I would like to say thank you to @jonnydossantos, for the suggestion of using numpy to replace pandas data frame. At that moment, the time consumption for each iteration run exceeded the limit too much. At the end, I used pertained model and simplified the notebook a lot, used tensor for dimensional matrix, avoid some conditional structures and did further selection of features.
\nLast but not least, the link of Notebook_NBEATS simplified version with limited features: https://www.kaggle.com/zhaoyuanyuanzyy/notebook-nbeats-with-simplified-features.
", "messages": [ { "id": 1725902, "postDate": "2022-03-17T13:58:22.297Z", "content": "Even though there are still two months to go, seeing I'm not a person in finance, I think for me is the moment to do a brief summary on my part and move on, I'm not found of checking lotto every two weeks.
\nFirst of all, I would like to say thank you to my teammate @rafalradwanski for his contribution of generation and selection of features. I cite here his partial work of features of LightGBM model after competition deadline (https://www.kaggle.com/rafalradwanski/g-research-features-to-increase-your-score/notebook), even though we've chosen two notebooks integrated with NBEATS models for final submission, this set of features for LightGBM model deserves a highlight.
\nRegarding two final submissions, I only talk about the simplified version with limited features (features widely shared among public notebooks, not enough with respect to quantity or quality), so I'm more likely to do a resume of what I've learnt after one month's concentrated work. Two main points, the selection of models and the struggle for final submission (especially the last one and a half weeks before deadline).
\nAs I mentioned, I'm not a quant trader, first time participating competition in the domain of finance, so I first read several other Kagglers' share of notebooks, fixed several potential good models, then I started the state of art. Due to the fact that I started quite late, the research period was extreme short. Our team's initial plan was to do practices on two models LightGBM and NBEATS, either the LightGBM or NBEATS has potential to reach a better performance, it depends on topics and how to use it, please refer to (The M5 Accuracy competition: Results, findings and conclusions: https://www.researchgate.net/publication/344487258_The_M5_Accuracy_competition_Results_findings_and_conclusions). Personally, I'm a technical person, NBEATS enables me to practice the technical 'routine' with not so bad result, so I focused on this model. Besides, I was amazed by the design of NBEATS, which motivated me to make more efforts on it (even while I encountered huge difficulties during submissions). Several gains after this practice, the configurations/parameters of training the model, the selection or balance of using data, and the procedure of dealing with the shape (flatten/squeeze or reshape/dimension).
\nAnother aspect I focused most at the late phase of the competition was conquering the submission errors. I would like to say thank you to @jonnydossantos, for the suggestion of using numpy to replace pandas data frame. At that moment, the time consumption for each iteration run exceeded the limit too much. At the end, I used pertained model and simplified the notebook a lot, used tensor for dimensional matrix, avoid some conditional structures and did further selection of features.
\nLast but not least, the link of Notebook_NBEATS simplified version with limited features: https://www.kaggle.com/zhaoyuanyuanzyy/notebook-nbeats-with-simplified-features.
", "rawMarkdown": "Even though there are still two months to go, seeing I'm not a person in finance, I think for me is the moment to do a brief summary on my part and move on, I'm not found of checking lotto every two weeks.\n\nFirst of all, I would like to say thank you to my teammate @rafalradwanski for his contribution of generation and selection of features. I cite here his partial work of features of LightGBM model after competition deadline (https://www.kaggle.com/rafalradwanski/g-research-features-to-increase-your-score/notebook), even though we've chosen two notebooks integrated with NBEATS models for final submission, this set of features for LightGBM model deserves a highlight.\n\nRegarding two final submissions, I only talk about the simplified version with limited features (features widely shared among public notebooks, not enough with respect to quantity or quality), so I'm more likely to do a resume of what I've learnt after one month's concentrated work. Two main points, the selection of models and the struggle for final submission (especially the last one and a half weeks before deadline). \n\nAs I mentioned, I'm not a quant trader, first time participating competition in the domain of finance, so I first read several other Kagglers' share of notebooks, fixed several potential good models, then I started the state of art. Due to the fact that I started quite late, the research period was extreme short. Our team's initial plan was to do practices on two models LightGBM and NBEATS, either the LightGBM or NBEATS has potential to reach a better performance, it depends on topics and how to use it, please refer to (The M5 Accuracy competition: Results, findings and conclusions: https://www.researchgate.net/publication/344487258_The_M5_Accuracy_competition_Results_findings_and_conclusions). Personally, I'm a technical person, NBEATS enables me to practice the technical 'routine' with not so bad result, so I focused on this model. Besides, I was amazed by the design of NBEATS, which motivated me to make more efforts on it (even while I encountered huge difficulties during submissions). Several gains after this practice, the configurations/parameters of training the model, the selection or balance of using data, and the procedure of dealing with the shape (flatten/squeeze or reshape/dimension).\n\nAnother aspect I focused most at the late phase of the competition was conquering the submission errors. I would like to say thank you to @jonnydossantos, for the suggestion of using numpy to replace pandas data frame. At that moment, the time consumption for each iteration run exceeded the limit too much. At the end, I used pertained model and simplified the notebook a lot, used tensor for dimensional matrix, avoid some conditional structures and did further selection of features.\n\nLast but not least, the link of Notebook_NBEATS simplified version with limited features: https://www.kaggle.com/zhaoyuanyuanzyy/notebook-nbeats-with-simplified-features.", "votes": 4 }, { "id": 1730961, "postDate": "2022-03-21T20:52:08.730Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1730961, "author_name": "", "author_url": "", "post_date": "2022-03-21T20:52:08.730000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1725902": "Even though there are still two months to go, seeing I'm not a person in finance, I think for me is the moment to do a brief summary on my part and move on, I'm not found of checking lotto every two weeks.\n\nFirst of all, I would like to say thank you to my teammate @rafalradwanski for his contribution of generation and selection of features. I cite here his partial work of features of LightGBM model after competition deadline (https://www.kaggle.com/rafalradwanski/g-research-features-to-increase-your-score/notebook), even though we've chosen two notebooks integrated with NBEATS models for final submission, this set of features for LightGBM model deserves a highlight.\n\nRegarding two final submissions, I only talk about the simplified version with limited features (features widely shared among public notebooks, not enough with respect to quantity or quality), so I'm more likely to do a resume of what I've learnt after one month's concentrated work. Two main points, the selection of models and the struggle for final submission (especially the last one and a half weeks before deadline). \n\nAs I mentioned, I'm not a quant trader, first time participating competition in the domain of finance, so I first read several other Kagglers' share of notebooks, fixed several potential good models, then I started the state of art. Due to the fact that I started quite late, the research period was extreme short. Our team's initial plan was to do practices on two models LightGBM and NBEATS, either the LightGBM or NBEATS has potential to reach a better performance, it depends on topics and how to use it, please refer to (The M5 Accuracy competition: Results, findings and conclusions: https://www.researchgate.net/publication/344487258_The_M5_Accuracy_competition_Results_findings_and_conclusions). Personally, I'm a technical person, NBEATS enables me to practice the technical 'routine' with not so bad result, so I focused on this model. Besides, I was amazed by the design of NBEATS, which motivated me to make more efforts on it (even while I encountered huge difficulties during submissions). Several gains after this practice, the configurations/parameters of training the model, the selection or balance of using data, and the procedure of dealing with the shape (flatten/squeeze or reshape/dimension).\n\nAnother aspect I focused most at the late phase of the competition was conquering the submission errors. I would like to say thank you to @jonnydossantos, for the suggestion of using numpy to replace pandas data frame. At that moment, the time consumption for each iteration run exceeded the limit too much. At the end, I used pertained model and simplified the notebook a lot, used tensor for dimensional matrix, avoid some conditional structures and did further selection of features.\n\nLast but not least, the link of Notebook_NBEATS simplified version with limited features: https://www.kaggle.com/zhaoyuanyuanzyy/notebook-nbeats-with-simplified-features.", "1730961": "" } }