I am stuck on this and maybe I am not thinking about it the right way.
\nHow can I encode the time series with only an ordinal number?
\nWhere can I read more about this? Because honestly, I don't get it.
I am stuck on this and maybe I am not thinking about it the right way.
\nHow can I encode the time series with only an ordinal number?
\nWhere can I read more about this? Because honestly, I don't get it.
What do you mean? date_id is already ordinal (1, 2, 3…)
", "rawMarkdown": "What do you mean? date_id is already ordinal (1, 2, 3...)", "votes": null }, { "id": "3055784", "postDate": "11/26/2024 06:00:49", "content": "Maybe I am thinking about it in the wrong way. If it is just an ordinal and it represents a day, how would i see repeating temporal patterns ?
", "rawMarkdown": "Maybe I am thinking about it in the wrong way. If it is just an ordinal and it represents a day, how would i see repeating temporal patterns ?", "votes": null }, { "id": "3056662", "postDate": "11/27/2024 08:52:12", "content": "I got inspiration from some nlp research，like bert
", "rawMarkdown": "I got inspiration from some nlp research，like bert", "votes": null }, { "id": "3059389", "postDate": "11/30/2024 16:21:42", "content": "You can feed a sequence of data into some models for relative time differences. In that case, the actual time is not relevant. It's just like words in a sentence. But perhaps different time_id's could also be a feature in the sense that prices tend to behave differently at different times, in which case you'd want to capture that. You could even do both, treating time as a feature and time as steps in a sequence.
", "rawMarkdown": "You can feed a sequence of data into some models for relative time differences. In that case, the actual time is not relevant. It's just like words in a sentence. But perhaps different time_id's could also be a feature in the sense that prices tend to behave differently at different times, in which case you'd want to capture that. You could even do both, treating time as a feature and time as steps in a sequence.", "votes": null }, { "id": "3059616", "postDate": "11/30/2024 22:59:33", "content": "You can try a sinusoidal embedding vector like they used for positional encodings in transformers, or even try mod 365 on the date_ids. Although I wouldn't think using date_ids would help given the non-stationary nature of the data.
", "rawMarkdown": "You can try a sinusoidal embedding vector like they used for positional encodings in transformers, or even try mod 365 on the date_ids. Although I wouldn't think using date_ids would help given the non-stationary nature of the data.", "votes": null } ], "comments": [ { "id": 3051519, "author_name": "natanlabarrere", "author_url": "", "post_date": "11/21/2024 11:14:22", "content": "What do you mean? date_id is already ordinal (1, 2, 3…)
", "votes": null, "replies": [ { "id": 3055784, "author_name": "jdsemrau", "author_url": "", "post_date": "11/26/2024 06:00:49", "content": "Maybe I am thinking about it in the wrong way. If it is just an ordinal and it represents a day, how would i see repeating temporal patterns ?
", "votes": null, "replies": [ { "id": 3059389, "author_name": "jimbeno", "author_url": "", "post_date": "11/30/2024 16:21:42", "content": "You can feed a sequence of data into some models for relative time differences. In that case, the actual time is not relevant. It's just like words in a sentence. But perhaps different time_id's could also be a feature in the sense that prices tend to behave differently at different times, in which case you'd want to capture that. You could even do both, treating time as a feature and time as steps in a sequence.
", "votes": null, "replies": [] }, { "id": 3059616, "author_name": "yanisfalaki", "author_url": "", "post_date": "11/30/2024 22:59:33", "content": "You can try a sinusoidal embedding vector like they used for positional encodings in transformers, or even try mod 365 on the date_ids. Although I wouldn't think using date_ids would help given the non-stationary nature of the data.
", "votes": null, "replies": [] } ] } ] }, { "id": 3056662, "author_name": "dearluna", "author_url": "", "post_date": "11/27/2024 08:52:12", "content": "I got inspiration from some nlp research，like bert
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3050656": "I am stuck on this and maybe I am not thinking about it the right way. \nHow can I encode the time series with only an ordinal number? \nWhere can I read more about this? Because honestly, I don't get it.", "3051519": "What do you mean? date_id is already ordinal (1, 2, 3...)", "3055784": "Maybe I am thinking about it in the wrong way. If it is just an ordinal and it represents a day, how would i see repeating temporal patterns ?", "3056662": "I got inspiration from some nlp research，like bert", "3059389": "You can feed a sequence of data into some models for relative time differences. In that case, the actual time is not relevant. It's just like words in a sentence. But perhaps different time_id's could also be a feature in the sense that prices tend to behave differently at different times, in which case you'd want to capture that. You could even do both, treating time as a feature and time as steps in a sequence.", "3059616": "You can try a sinusoidal embedding vector like they used for positional encodings in transformers, or even try mod 365 on the date_ids. Although I wouldn't think using date_ids would help given the non-stationary nature of the data." }, "source": "meta" }