Hello, everyone! How are you?
\nBased on your experiences, what does this competition need in order to achieve good results?
\nIf I understand correctly, feature engineering doesn't seem very important here. So, what is most important in this case?
What do you recommend? This is my first competition with a prize on the line.
", "messages": [ { "id": "3026972", "postDate": "10/24/2024 11:56:29", "content": "Hello, everyone! How are you?
\nBased on your experiences, what does this competition need in order to achieve good results?
\nIf I understand correctly, feature engineering doesn't seem very important here. So, what is most important in this case?
What do you recommend? This is my first competition with a prize on the line.
", "rawMarkdown": "Hello, everyone! How are you?\n\nBased on your experiences, what does this competition need in order to achieve good results?\nIf I understand correctly, feature engineering doesn't seem very important here. So, what is most important in this case?\n\n- AutoRegressive models?\n- Just hyperparameter tuning?\n- Feature reduction?\n- Ensemble methods?\n\nWhat do you recommend? This is my first competition with a prize on the line.", "votes": null }, { "id": "3027166", "postDate": "10/24/2024 14:44:11", "content": "@josviniciussa How about using time series neural networks like LSTM or transformer?
", "rawMarkdown": "josviniciussa How about using time series neural networks like LSTM or transformer?", "votes": null }, { "id": "3027192", "postDate": "10/24/2024 14:59:51", "content": "I used GRU for time-series modelling and the performance is not good (yet).
", "rawMarkdown": "I used GRU for time-series modelling and the performance is not good (yet).", "votes": null }, { "id": "3027893", "postDate": "10/25/2024 12:02:33", "content": "Excuse me, but how did you reach the conclusion that feature engineering isn't very important here? I'm only asking because I'm concerned that I might be missing something or misunderstanding a key aspect of the competition. Otherwise, I would argue that feature engineering is one of the most critical aspects of this competition, as it is in any time series competition. Top solutions typically use GBMs, which lack the ability to handle temporal dependencies unless features are added to give the model a sense of time.
\nPlease correct me if I'm because I don't which to spend much time walking in the wrong direction
\n\nWhen approaching modeling problems in modern financial markets, there are many reasons to believe that the problems you are trying to solve are impossible. Even if you put aside the beliefs that the prices of financial instruments rationally reflect all available information, you’ll have to grapple with time series and distributions that have properties you don’t encounter in other sorts of modeling problems. Distributions can be famously fat-tailed, time series can be non-stationary, and data can generally fail to satisfy a lot of the underlying assumptions on which very successful statistical approaches rely.
\n
So, I think FE is the one of the most important steps.
", "rawMarkdown": ">When approaching modeling problems in modern financial markets, there are many reasons to believe that the problems you are trying to solve are impossible. Even if you put aside the beliefs that the prices of financial instruments rationally reflect all available information, you’ll have to grapple with time series and distributions that have properties you don’t encounter in other sorts of modeling problems. Distributions can be famously fat-tailed, time series can be non-stationary, and data can generally fail to satisfy a lot of the underlying assumptions on which very successful statistical approaches rely.\n\nSo, I think FE is the one of the most important steps.", "votes": null }, { "id": "3028181", "postDate": "10/25/2024 17:28:17", "content": "My main reason for thinking that feature engineering is unlikely in this case is that we don’t know the nature of each feature. So, how could we perform effective feature engineering? For instance, how can we combine features without knowing the labels or meanings of each feature?
", "rawMarkdown": "My main reason for thinking that feature engineering is unlikely in this case is that we don’t know the nature of each feature. So, how could we perform effective feature engineering? For instance, how can we combine features without knowing the labels or meanings of each feature?", "votes": null } ], "comments": [ { "id": 3027166, "author_name": "chumajin", "author_url": "", "post_date": "10/24/2024 14:44:11", "content": "@josviniciussa How about using time series neural networks like LSTM or transformer?
", "votes": null, "replies": [ { "id": 3027192, "author_name": "shiyili", "author_url": "", "post_date": "10/24/2024 14:59:51", "content": "I used GRU for time-series modelling and the performance is not good (yet).
", "votes": null, "replies": [] } ] }, { "id": 3027893, "author_name": "aymanallawi", "author_url": "", "post_date": "10/25/2024 12:02:33", "content": "Excuse me, but how did you reach the conclusion that feature engineering isn't very important here? I'm only asking because I'm concerned that I might be missing something or misunderstanding a key aspect of the competition. Otherwise, I would argue that feature engineering is one of the most critical aspects of this competition, as it is in any time series competition. Top solutions typically use GBMs, which lack the ability to handle temporal dependencies unless features are added to give the model a sense of time.
\nPlease correct me if I'm because I don't which to spend much time walking in the wrong direction
\n\nWhen approaching modeling problems in modern financial markets, there are many reasons to believe that the problems you are trying to solve are impossible. Even if you put aside the beliefs that the prices of financial instruments rationally reflect all available information, you’ll have to grapple with time series and distributions that have properties you don’t encounter in other sorts of modeling problems. Distributions can be famously fat-tailed, time series can be non-stationary, and data can generally fail to satisfy a lot of the underlying assumptions on which very successful statistical approaches rely.
\n
So, I think FE is the one of the most important steps.
", "votes": null, "replies": [ { "id": 3028181, "author_name": "josviniciussa", "author_url": "", "post_date": "10/25/2024 17:28:17", "content": "My main reason for thinking that feature engineering is unlikely in this case is that we don’t know the nature of each feature. So, how could we perform effective feature engineering? For instance, how can we combine features without knowing the labels or meanings of each feature?
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3026972": "Hello, everyone! How are you?\n\nBased on your experiences, what does this competition need in order to achieve good results?\nIf I understand correctly, feature engineering doesn't seem very important here. So, what is most important in this case?\n\n- AutoRegressive models?\n- Just hyperparameter tuning?\n- Feature reduction?\n- Ensemble methods?\n\nWhat do you recommend? This is my first competition with a prize on the line.", "3027166": "josviniciussa How about using time series neural networks like LSTM or transformer?", "3027192": "I used GRU for time-series modelling and the performance is not good (yet).", "3027893": "Excuse me, but how did you reach the conclusion that feature engineering isn't very important here? I'm only asking because I'm concerned that I might be missing something or misunderstanding a key aspect of the competition. Otherwise, I would argue that feature engineering is one of the most critical aspects of this competition, as it is in any time series competition. Top solutions typically use GBMs, which lack the ability to handle temporal dependencies unless features are added to give the model a sense of time.\nPlease correct me if I'm because I don't which to spend much time walking in the wrong direction", "3028122": ">When approaching modeling problems in modern financial markets, there are many reasons to believe that the problems you are trying to solve are impossible. Even if you put aside the beliefs that the prices of financial instruments rationally reflect all available information, you’ll have to grapple with time series and distributions that have properties you don’t encounter in other sorts of modeling problems. Distributions can be famously fat-tailed, time series can be non-stationary, and data can generally fail to satisfy a lot of the underlying assumptions on which very successful statistical approaches rely.\n\nSo, I think FE is the one of the most important steps.", "3028181": "My main reason for thinking that feature engineering is unlikely in this case is that we don’t know the nature of each feature. So, how could we perform effective feature engineering? For instance, how can we combine features without knowing the labels or meanings of each feature?" }, "source": "meta" }