I understand that this competition may have masked the original data to prevent data leakage. However, it seems to have turned into a competition that relatively tests feature engineering and parameter tuning, and I am not sure if I am correct in this regard.
\nI also question the practical value of this model because the market is a place that constantly evolves with feedback, and simple regression may not hold high commercial value.
\nOf course, it is also possible that I have just started learning about machine learning and my understanding is still not deep enough.
", "messages": [ { "id": "3208343", "postDate": "05/23/2025 23:57:29", "content": "I understand that this competition may have masked the original data to prevent data leakage. However, it seems to have turned into a competition that relatively tests feature engineering and parameter tuning, and I am not sure if I am correct in this regard.
\nI also question the practical value of this model because the market is a place that constantly evolves with feedback, and simple regression may not hold high commercial value.
\nOf course, it is also possible that I have just started learning about machine learning and my understanding is still not deep enough.
", "rawMarkdown": "I understand that this competition may have masked the original data to prevent data leakage. However, it seems to have turned into a competition that relatively tests feature engineering and parameter tuning, and I am not sure if I am correct in this regard. \n\nI also question the practical value of this model because the market is a place that constantly evolves with feedback, and simple regression may not hold high commercial value. \n\nOf course, it is also possible that I have just started learning about machine learning and my understanding is still not deep enough.", "votes": null }, { "id": "3208543", "postDate": "05/24/2025 10:10:35", "content": "You're absolutely right to raise those points — and I think many participants are feeling the same way.
\nThe masking and shuffling of timestamps definitely limit the use of temporal patterns, which shifts the focus more toward pure feature engineering and model tuning. It can feel a bit artificial, but I guess the goal is to create a level playing field and prevent any leakage from time-based structures.
\nRegarding data leakage — for example, you can identify the feature most correlated with the target, take its value from the future, and then achieve very high scores on the Kaggle test set.
\nBut this has absolutely no real-world value, because in a real scenario, you simply wouldn’t have access to those future values when making predictions.
", "rawMarkdown": "You're absolutely right to raise those points — and I think many participants are feeling the same way.\n\nThe masking and shuffling of timestamps definitely limit the use of temporal patterns, which shifts the focus more toward pure feature engineering and model tuning. It can feel a bit artificial, but I guess the goal is to create a level playing field and prevent any leakage from time-based structures.\n\nRegarding data leakage — for example, you can identify the feature most correlated with the target, take its value from the future, and then achieve very high scores on the Kaggle test set.\n\nBut this has absolutely no real-world value, because in a real scenario, you simply wouldn’t have access to those future values when making predictions.", "votes": null } ], "comments": [ { "id": 3208543, "author_name": "michalinahulak", "author_url": "", "post_date": "05/24/2025 10:10:35", "content": "You're absolutely right to raise those points — and I think many participants are feeling the same way.
\nThe masking and shuffling of timestamps definitely limit the use of temporal patterns, which shifts the focus more toward pure feature engineering and model tuning. It can feel a bit artificial, but I guess the goal is to create a level playing field and prevent any leakage from time-based structures.
\nRegarding data leakage — for example, you can identify the feature most correlated with the target, take its value from the future, and then achieve very high scores on the Kaggle test set.
\nBut this has absolutely no real-world value, because in a real scenario, you simply wouldn’t have access to those future values when making predictions.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3208343": "I understand that this competition may have masked the original data to prevent data leakage. However, it seems to have turned into a competition that relatively tests feature engineering and parameter tuning, and I am not sure if I am correct in this regard. \n\nI also question the practical value of this model because the market is a place that constantly evolves with feedback, and simple regression may not hold high commercial value. \n\nOf course, it is also possible that I have just started learning about machine learning and my understanding is still not deep enough.", "3208543": "You're absolutely right to raise those points — and I think many participants are feeling the same way.\n\nThe masking and shuffling of timestamps definitely limit the use of temporal patterns, which shifts the focus more toward pure feature engineering and model tuning. It can feel a bit artificial, but I guess the goal is to create a level playing field and prevent any leakage from time-based structures.\n\nRegarding data leakage — for example, you can identify the feature most correlated with the target, take its value from the future, and then achieve very high scores on the Kaggle test set.\n\nBut this has absolutely no real-world value, because in a real scenario, you simply wouldn’t have access to those future values when making predictions." }, "source": "meta" }