I self had an idea of RL, would have been a challenge and interesting journey but took the path of LGBM/XGB + TA features. What high level approach did you chose to launch in the final selection?
\nI can summarize the answer here if we get a couple of answers, maybe can be fun information to aggregate…
\nArchitecture:
\nGBM
\nLSTM w/ secondary losses
\nOnline Learning
Features:
\nAs this was a more feature-based competition with extensive feature creation solutions, it's in place to list them here as well, from general findings to single features that stands out from the crowd. Below I summarize what have been posted.
\nI self had an idea of RL, would have been a challenge and interesting journey but took the path of LGBM/XGB + TA features. What high level approach did you chose to launch in the final selection?
\nI can summarize the answer here if we get a couple of answers, maybe can be fun information to aggregate…
\nArchitecture:
\nGBM
\nLSTM w/ secondary losses
\nOnline Learning
Features:
\nAs this was a more feature-based competition with extensive feature creation solutions, it's in place to list them here as well, from general findings to single features that stands out from the crowd. Below I summarize what have been posted.
\nI have tried some advanced tabular NN models from latest papers but it seems they are not outperforming GBM. Features are the most important part in this competition.
", "rawMarkdown": "I have tried some advanced tabular NN models from latest papers but it seems they are not outperforming GBM. Features are the most important part in this competition.", "votes": null }, { "id": "1673179", "postDate": "02/02/2022 14:59:24", "content": "Yes tested some other models/approaches as well, they also needed to manage the time limit, LGBM with or without gpu and xgb with gpu are fast. Indeed, 99% of the time went to the feature/data creations :)
", "rawMarkdown": "Yes tested some other models/approaches as well, they also needed to manage the time limit, LGBM with or without gpu and xgb with gpu are fast. Indeed, 99% of the time went to the feature/data creations :)", "votes": null }, { "id": "1673186", "postDate": "02/02/2022 15:04:31", "content": "On line learning. Very good results locally (near 0.06 on the local replication of the LB). But I haven't been able to get matching results on LB. Possible leakage. Probable leakage.
", "rawMarkdown": "On line learning. Very good results locally (near 0.06 on the local replication of the LB). But I haven't been able to get matching results on LB. Possible leakage. Probable leakage.", "votes": null }, { "id": "1673191", "postDate": "02/02/2022 15:08:55", "content": "I tried many things feature wise. Nothing really improving scores above what I already shared in my on-line feature engineering notebook. Maybe correlation features for an inconsistent 0.001.
", "rawMarkdown": "I tried many things feature wise. Nothing really improving scores above what I already shared in my on-line feature engineering notebook. Maybe correlation features for an inconsistent 0.001.", "votes": null }, { "id": "1673212", "postDate": "02/02/2022 15:22:25", "content": "Interesting, online learning, didn't try that, good one!
", "rawMarkdown": "Interesting, online learning, didn't try that, good one!", "votes": null }, { "id": "1673342", "postDate": "02/02/2022 17:02:37", "content": "Advanced FE + 14 different LGBMRegressor models with tunning + Kaggle dataset with pre-trained models…. last notebook ran 40 sec.
", "rawMarkdown": "Advanced FE + 14 different LGBMRegressor models with tunning + Kaggle dataset with pre-trained models.... last notebook ran 40 sec.", "votes": null }, { "id": "1673455", "postDate": "02/02/2022 18:31:56", "content": "@gogo827jz I'm curious to know what was your most important feature.
\nMine was lag of ( close - open)/open
I made 3 fold (not-countinous) LGBM with rolling aggregations and some sort of lagged target. The most important feature for me was mean of past returns ( lag1 / lag15) for all assets and then second was hull moving average.
", "rawMarkdown": "I made 3 fold (not-countinous) LGBM with rolling aggregations and some sort of lagged target. The most important feature for me was mean of past returns ( lag1 / lag15) for all assets and then second was hull moving average.", "votes": null }, { "id": "1673624", "postDate": "02/02/2022 21:04:36", "content": "I used a multibranch lstm model with secondary losses to predict the market return and the individual returns of the coins.
", "rawMarkdown": "I used a multibranch lstm model with secondary losses to predict the market return and the individual returns of the coins.", "votes": null }, { "id": "1673749", "postDate": "02/02/2022 23:55:46", "content": "Nice with a different approach and it sounds like a good model creation!
", "rawMarkdown": "Nice with a different approach and it sounds like a good model creation!", "votes": null }, { "id": "1674804", "postDate": "02/03/2022 19:44:27", "content": "my gold features: residualization. Residualization is the technique used to calculate the target but I used it to some features. I used this method for return related features it boosted my local from 0.04 to 0.06 (with just 5 more features added)
", "rawMarkdown": "my gold features: residualization. Residualization is the technique used to calculate the target but I used it to some features. I used this method for return related features it boosted my local from 0.04 to 0.06 (with just 5 more features added)", "votes": null }, { "id": "1675442", "postDate": "02/04/2022 09:18:45", "content": "Good finding! Hope it holds in the final runs, that's a good score.
", "rawMarkdown": "Good finding! Hope it holds in the final runs, that's a good score.", "votes": null }, { "id": "1675680", "postDate": "02/04/2022 12:09:03", "content": "I split the recomputed target into 16 chunks at 16min interval then run arima(1,0,0) to predict t0 from t(-i*16).
\nI got a better no leakage score (0.027) on lb than with lgbm.
\nRationale was past targets include info from important features (close, beta) and autocorrelation present at t-16.
I have tried some advanced tabular NN models from latest papers but it seems they are not outperforming GBM. Features are the most important part in this competition.
", "votes": null, "replies": [ { "id": 1673179, "author_name": "kirderf", "author_url": "", "post_date": "02/02/2022 14:59:24", "content": "Yes tested some other models/approaches as well, they also needed to manage the time limit, LGBM with or without gpu and xgb with gpu are fast. Indeed, 99% of the time went to the feature/data creations :)
", "votes": null, "replies": [] }, { "id": 1673191, "author_name": "lucasmorin", "author_url": "", "post_date": "02/02/2022 15:08:55", "content": "I tried many things feature wise. Nothing really improving scores above what I already shared in my on-line feature engineering notebook. Maybe correlation features for an inconsistent 0.001.
", "votes": null, "replies": [] }, { "id": 1673455, "author_name": "amedprof", "author_url": "", "post_date": "02/02/2022 18:31:56", "content": "@gogo827jz I'm curious to know what was your most important feature.
\nMine was lag of ( close - open)/open
On line learning. Very good results locally (near 0.06 on the local replication of the LB). But I haven't been able to get matching results on LB. Possible leakage. Probable leakage.
", "votes": null, "replies": [ { "id": 1673212, "author_name": "kirderf", "author_url": "", "post_date": "02/02/2022 15:22:25", "content": "Interesting, online learning, didn't try that, good one!
", "votes": null, "replies": [] } ] }, { "id": 1673342, "author_name": "vbmokin", "author_url": "", "post_date": "02/02/2022 17:02:37", "content": "Advanced FE + 14 different LGBMRegressor models with tunning + Kaggle dataset with pre-trained models…. last notebook ran 40 sec.
", "votes": null, "replies": [] }, { "id": 1673603, "author_name": "bturan19", "author_url": "", "post_date": "02/02/2022 20:38:05", "content": "I made 3 fold (not-countinous) LGBM with rolling aggregations and some sort of lagged target. The most important feature for me was mean of past returns ( lag1 / lag15) for all assets and then second was hull moving average.
", "votes": null, "replies": [] }, { "id": 1673624, "author_name": "petersk20", "author_url": "", "post_date": "02/02/2022 21:04:36", "content": "I used a multibranch lstm model with secondary losses to predict the market return and the individual returns of the coins.
", "votes": null, "replies": [ { "id": 1673749, "author_name": "kirderf", "author_url": "", "post_date": "02/02/2022 23:55:46", "content": "Nice with a different approach and it sounds like a good model creation!
", "votes": null, "replies": [] } ] }, { "id": 1674804, "author_name": "axzhang", "author_url": "", "post_date": "02/03/2022 19:44:27", "content": "my gold features: residualization. Residualization is the technique used to calculate the target but I used it to some features. I used this method for return related features it boosted my local from 0.04 to 0.06 (with just 5 more features added)
", "votes": null, "replies": [ { "id": 1675442, "author_name": "kirderf", "author_url": "", "post_date": "02/04/2022 09:18:45", "content": "Good finding! Hope it holds in the final runs, that's a good score.
", "votes": null, "replies": [] } ] }, { "id": 1675680, "author_name": "mariusbadea", "author_url": "", "post_date": "02/04/2022 12:09:03", "content": "I split the recomputed target into 16 chunks at 16min interval then run arima(1,0,0) to predict t0 from t(-i*16).
\nI got a better no leakage score (0.027) on lb than with lgbm.
\nRationale was past targets include info from important features (close, beta) and autocorrelation present at t-16.