In most the of the time-series forecasting challenges we typically use Boosting tree based models with rolling window features (to incorporate time-series aspect in the model). Can we use such features in this challenge as in the API (also in the final evaluation phase) we only have access to data from one timestamp. Can someone please elaborate on this and if possible point me to some code snippet or notebook, if such features are already implemented.
\nThanks a lot in advance !!
", "messages": [ { "id": "1636244", "postDate": "01/02/2022 17:45:31", "content": "In most the of the time-series forecasting challenges we typically use Boosting tree based models with rolling window features (to incorporate time-series aspect in the model). Can we use such features in this challenge as in the API (also in the final evaluation phase) we only have access to data from one timestamp. Can someone please elaborate on this and if possible point me to some code snippet or notebook, if such features are already implemented.
\nThanks a lot in advance !!
", "rawMarkdown": "In most the of the time-series forecasting challenges we typically use Boosting tree based models with rolling window features (to incorporate time-series aspect in the model). Can we use such features in this challenge as in the API (also in the final evaluation phase) we only have access to data from one timestamp. Can someone please elaborate on this and if possible point me to some code snippet or notebook, if such features are already implemented. \n\nThanks a lot in advance !!", "votes": null }, { "id": "1636307", "postDate": "01/02/2022 18:42:07", "content": "Hello, @bikash02 I think is possible, you need to modify the inference function to accumulate data and generate your rolling features. I will release a notebook today with this functionality.
", "rawMarkdown": "Hello, @bikash02 I think is possible, you need to modify the inference function to accumulate data and generate your rolling features. I will release a notebook today with this functionality.", "votes": null }, { "id": "1636406", "postDate": "01/02/2022 20:55:49", "content": "I showed how to do this efficiently (300 features in less than 10 minutes of time submission) using numpy in this notebook: https://www.kaggle.com/lucasmorin/on-line-feature-engineering.
", "rawMarkdown": "I showed how to do this efficiently (300 features in less than 10 minutes of time submission) using numpy in this notebook: https://www.kaggle.com/lucasmorin/on-line-feature-engineering.", "votes": null }, { "id": "1636427", "postDate": "01/02/2022 21:41:33", "content": "Great, I will check it out. Thanks
", "rawMarkdown": "Great, I will check it out. Thanks", "votes": null }, { "id": "1637867", "postDate": "01/04/2022 09:12:39", "content": "I suggest the following discussion.
\nhttps://www.kaggle.com/c/g-research-crypto-forecasting/discussion/285103#1569878
I think the difficult part is you need to store your test data features for predicting new time ticks. Using time series features is really hard for predicting.
", "rawMarkdown": "I think the difficult part is you need to store your test data features for predicting new time ticks. Using time series features is really hard for predicting.", "votes": null }, { "id": "1638821", "postDate": "01/05/2022 04:31:09", "content": "Hello @lucasmorin, I will look at your implementation to get some ideas.
", "rawMarkdown": "Hello @lucasmorin, I will look at your implementation to get some ideas.", "votes": null } ], "comments": [ { "id": 1636307, "author_name": "cv13j0", "author_url": "", "post_date": "01/02/2022 18:42:07", "content": "Hello, @bikash02 I think is possible, you need to modify the inference function to accumulate data and generate your rolling features. I will release a notebook today with this functionality.
", "votes": null, "replies": [ { "id": 1636427, "author_name": "bikash02", "author_url": "", "post_date": "01/02/2022 21:41:33", "content": "Great, I will check it out. Thanks
", "votes": null, "replies": [] }, { "id": 1638206, "author_name": "xiejiancong", "author_url": "", "post_date": "01/04/2022 14:30:11", "content": "I think the difficult part is you need to store your test data features for predicting new time ticks. Using time series features is really hard for predicting.
", "votes": null, "replies": [] } ] }, { "id": 1636406, "author_name": "lucasmorin", "author_url": "", "post_date": "01/02/2022 20:55:49", "content": "I showed how to do this efficiently (300 features in less than 10 minutes of time submission) using numpy in this notebook: https://www.kaggle.com/lucasmorin/on-line-feature-engineering.
", "votes": null, "replies": [ { "id": 1638821, "author_name": "cv13j0", "author_url": "", "post_date": "01/05/2022 04:31:09", "content": "Hello @lucasmorin, I will look at your implementation to get some ideas.
", "votes": null, "replies": [] } ] }, { "id": 1637867, "author_name": "joeonaka", "author_url": "", "post_date": "01/04/2022 09:12:39", "content": "I suggest the following discussion.
\nhttps://www.kaggle.com/c/g-research-crypto-forecasting/discussion/285103#1569878