The train and the test overlap, that's why there are such high scores, but if if you only use Train until 2021-06-13, what is your best score?
\nFor now I am not full with the models I am with feature engineering , but some experiment without model like this https://www.kaggle.com/lucasmorin/minimal-fast-baseline
\nI got score 0.034
", "messages": [ { "id": 1579160, "postDate": "2021-11-11T15:41:13.110Z", "content": "The train and the test overlap, that's why there are such high scores, but if if you only use Train until 2021-06-13, what is your best score?
\nFor now I am not full with the models I am with feature engineering , but some experiment without model like this https://www.kaggle.com/lucasmorin/minimal-fast-baseline
\nI got score 0.034
", "rawMarkdown": "The train and the test overlap, that's why there are such high scores, but if if you only use Train until 2021-06-13, what is your best score?\n\nFor now I am not full with the models I am with feature engineering , but some experiment without model like this https://www.kaggle.com/lucasmorin/minimal-fast-baseline\n\nI got score 0.034", "votes": 20 }, { "id": 1579190, "postDate": "2021-11-11T16:07:38.280Z", "content": "Well done - mine is 0.0195 = ridge regression with some feature engineering (vwap and volume-based). Would you mind sharing any details about yours?
", "rawMarkdown": "Well done - mine is 0.0195 = ridge regression with some feature engineering (vwap and volume-based). Would you mind sharing any details about yours?", "votes": 8 }, { "id": 1581720, "postDate": "2021-11-14T06:58:34.270Z", "content": "mine is 0.0193 = lgbm regression with some feature engineering
", "rawMarkdown": "mine is 0.0193 = lgbm regression with some feature engineering", "votes": 3 }, { "id": 1582729, "postDate": "2021-11-15T07:55:49.527Z", "content": "Mine is 0.024 , Ridge with feature engineering
", "rawMarkdown": "Mine is 0.024 , Ridge with feature engineering\n", "votes": 1 }, { "id": 1579366, "postDate": "2021-11-11T20:05:11.797Z", "content": "Slightly above 3% on local CV.
", "rawMarkdown": "Slightly above 3% on local CV.", "votes": 2, "replies": [ { "id": 1579377, "postDate": "2021-11-11T20:23:56.970Z", "content": "CV can be misleading for time series because it intertwines the training/validation rows. Instead, a better split would be a simple train test split where everything before a certain date is training and everything after is testing. This is more realistic because when we are predicting cryptocurrency returns we will not have training instances of the future.
", "rawMarkdown": "CV can be misleading for time series because it intertwines the training/validation rows. Instead, a better split would be a simple train test split where everything before a certain date is training and everything after is testing. This is more realistic because when we are predicting cryptocurrency returns we will not have training instances of the future.", "votes": 2 }, { "id": 1579395, "postDate": "2021-11-11T20:46:58.943Z", "content": "I use something like this:
\nX - train_0 - X - test_0 - X - train_1 - X - test_1 - …. where X are 1 month embargos (also used for building lagged features).
", "rawMarkdown": "I use something like this:\n\nX - train_0 - X - test_0 - X - train_1 - X - test_1 - .... where X are 1 month embargos (also used for building lagged features). ", "votes": 1 } ] }, { "id": 1580800, "postDate": "2021-11-13T06:01:25.987Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1579190, "author_name": "Konrad Banachewicz", "author_url": "", "post_date": "2021-11-11T16:07:38.280000", "content": "Well done - mine is 0.0195 = ridge regression with some feature engineering (vwap and volume-based). Would you mind sharing any details about yours?
", "votes": 8, "replies": [] }, { "id": 1581720, "author_name": "Aleksandr Larko", "author_url": "", "post_date": "2021-11-14T06:58:34.270000", "content": "mine is 0.0193 = lgbm regression with some feature engineering
", "votes": 3, "replies": [] }, { "id": 1582729, "author_name": "Saumil Agrawal", "author_url": "", "post_date": "2021-11-15T07:55:49.527000", "content": "Mine is 0.024 , Ridge with feature engineering
", "votes": 1, "replies": [] }, { "id": 1579366, "author_name": "Lucas Morin", "author_url": "", "post_date": "2021-11-11T20:05:11.797000", "content": "Slightly above 3% on local CV.
", "votes": 2, "replies": [ { "id": 1579377, "author_name": "Kyle Peters", "author_url": "", "post_date": "2021-11-11T20:23:56.970000", "content": "CV can be misleading for time series because it intertwines the training/validation rows. Instead, a better split would be a simple train test split where everything before a certain date is training and everything after is testing. This is more realistic because when we are predicting cryptocurrency returns we will not have training instances of the future.
", "votes": 2, "replies": [] }, { "id": 1579395, "author_name": "Lucas Morin", "author_url": "", "post_date": "2021-11-11T20:46:58.943000", "content": "I use something like this:
\nX - train_0 - X - test_0 - X - train_1 - X - test_1 - …. where X are 1 month embargos (also used for building lagged features).
", "votes": 1, "replies": [] } ] }, { "id": 1580800, "author_name": "", "author_url": "", "post_date": "2021-11-13T06:01:25.987000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1579160": "The train and the test overlap, that's why there are such high scores, but if if you only use Train until 2021-06-13, what is your best score?\n\nFor now I am not full with the models I am with feature engineering , but some experiment without model like this https://www.kaggle.com/lucasmorin/minimal-fast-baseline\n\nI got score 0.034", "1579190": "Well done - mine is 0.0195 = ridge regression with some feature engineering (vwap and volume-based). Would you mind sharing any details about yours?", "1581720": "mine is 0.0193 = lgbm regression with some feature engineering", "1582729": "Mine is 0.024 , Ridge with feature engineering\n", "1579366": "Slightly above 3% on local CV.", "1580800": "" } }