Congratulations to the winners, to all the players and thanks to Kaggle and all the organizers for this tabular actuarial challenge ! And thank you to those who shared some tricks during the challenge !
\nMy best CV result was a blend of Catboost, LGBM and NN (and no extratree despite of my notebook published in january). I had a worse CV with a single NN than a single CatBoost or a single LGBM. But NN model in blend improve my CV (and not extratree model, sorry !)
\nI didn't trust public leaderboard, and I took many times to build robust pipelines (without target leakage) and a clean CV.
\nI had only one submission with a RMSE under 30 000 on the public LB ; it was my best CV.
\nI'm proud to be the only person in the 20 first places to have less than 10 submissions on public leaderboard.
\nI'm glad to have tried a NN on tabular data, I've learned a lot about text mining and how to describe data with seaborn plots.
Hope to see you soon in an other challenge !
\nAlexandre
Hey Alexandre, I am very happy to see that my work help you.
\nCongratulations for your amazing performance !
Thank you for sharing!
\nAmazing under 10 submissions! Am very interested in how you made your local CV pipeline that robust!
Especially so since I'm guessing your 2nd magic feature is very prone to target leakage right?
", "votes": 1, "replies": [] }, { "id": 1274244, "author_name": "aldparis", "author_url": "", "post_date": "2021-04-15T06:15:59.010000", "content": "Thank you for your comment and your question.
\n\n\nI'm guessing your 2nd magic feature is very prone to target leakage
\n
I don't think so, but maybe I'm wrong.
\nI have created the ClaimDescription score like a way of data reduction : instead of a text, or instead of many one hot text features, i had only a continuous feature with this score, usefull.
\nThis ClaimDescription score was build separately for each fold only on training sets during CV in a pipeline of transformations, and before LGBM or catboost or NN trainings.
\nThere would be target leakage if i had trained this ClaimDescription score on training sets and on valid sets before training LGBM, NN etc. in CV , but I didn't.
\nAnd those transformers, transformed later valid samples (and test samples of course) just before predictions. Those valid and test samples were not used during training of transformers.
Also, if CV RMSE (RMSE on valid sets) is lower on my final LGBM when I'm using the ClaimDescriptionScore feature, than when I'm not, I'm almost certain that this ClaimDescription Score will improve the LGBM RMSE on the PRIVATE leaderboard.
\nNever mind if the RMSE is higher on PUBLIC leaderboard ; anyway, I didn't submit this little LGBM difference on public LB.
Since the beginning of the challenge, I've made all my decisions by this way (number of topics with LDA implementation etc.). That's why I had submitted so few. At the end of the challenge, all my new CV RMSE on new ideas told me that those new ideas shouldn't improve my RMSE on unseen private leaderboard, so I didn't submitted them.
\nThe notebook published during january is like that.
Thanks to scikit-learn for transformer classes and pipeline, there are very powerfull tools to make decisions : a link.
", "votes": 2, "replies": [] }, { "id": 1274310, "author_name": "teehoong", "author_url": "", "post_date": "2021-04-15T07:20:53.147000", "content": "Thank you Ald. 👍
", "votes": 1, "replies": [] } ] }, { "id": 1275331, "author_name": "Abhishek252", "author_url": "", "post_date": "2021-04-16T07:26:55.110000", "content": "amazingly done buddy!
", "votes": 0, "replies": [] } ], "comment_count": 6, "messages": [ { "id": 1274001, "postDate": "2021-04-14T21:47:32.963Z", "content": "Congratulations to the winners, to all the players and thanks to Kaggle and all the organizers for this tabular actuarial challenge ! And thank you to those who shared some tricks during the challenge !
\nMy best CV result was a blend of Catboost, LGBM and NN (and no extratree despite of my notebook published in january). I had a worse CV with a single NN than a single CatBoost or a single LGBM. But NN model in blend improve my CV (and not extratree model, sorry !)
\nI didn't trust public leaderboard, and I took many times to build robust pipelines (without target leakage) and a clean CV.
\nI had only one submission with a RMSE under 30 000 on the public LB ; it was my best CV.
\nI'm proud to be the only person in the 20 first places to have less than 10 submissions on public leaderboard.
\nI'm glad to have tried a NN on tabular data, I've learned a lot about text mining and how to describe data with seaborn plots.
Hope to see you soon in an other challenge !
\nAlexandre
Hey Alexandre, I am very happy to see that my work help you.
\nCongratulations for your amazing performance !
Thank you for sharing!
\nAmazing under 10 submissions! Am very interested in how you made your local CV pipeline that robust!
Especially so since I'm guessing your 2nd magic feature is very prone to target leakage right?
", "rawMarkdown": "Especially so since I'm guessing your 2nd magic feature is very prone to target leakage right?", "votes": 1 }, { "id": 1274244, "postDate": "2021-04-15T06:15:59.010Z", "content": "Thank you for your comment and your question.
\n\n\nI'm guessing your 2nd magic feature is very prone to target leakage
\n
I don't think so, but maybe I'm wrong.
\nI have created the ClaimDescription score like a way of data reduction : instead of a text, or instead of many one hot text features, i had only a continuous feature with this score, usefull.
\nThis ClaimDescription score was build separately for each fold only on training sets during CV in a pipeline of transformations, and before LGBM or catboost or NN trainings.
\nThere would be target leakage if i had trained this ClaimDescription score on training sets and on valid sets before training LGBM, NN etc. in CV , but I didn't.
\nAnd those transformers, transformed later valid samples (and test samples of course) just before predictions. Those valid and test samples were not used during training of transformers.
Also, if CV RMSE (RMSE on valid sets) is lower on my final LGBM when I'm using the ClaimDescriptionScore feature, than when I'm not, I'm almost certain that this ClaimDescription Score will improve the LGBM RMSE on the PRIVATE leaderboard.
\nNever mind if the RMSE is higher on PUBLIC leaderboard ; anyway, I didn't submit this little LGBM difference on public LB.
Since the beginning of the challenge, I've made all my decisions by this way (number of topics with LDA implementation etc.). That's why I had submitted so few. At the end of the challenge, all my new CV RMSE on new ideas told me that those new ideas shouldn't improve my RMSE on unseen private leaderboard, so I didn't submitted them.
\nThe notebook published during january is like that.
Thanks to scikit-learn for transformer classes and pipeline, there are very powerfull tools to make decisions : a link.
", "rawMarkdown": "Thank you for your comment and your question.\n>I'm guessing your 2nd magic feature is very prone to target leakage\n\nI don't think so, but maybe I'm wrong.\n\nI have created the ClaimDescription score like a way of data reduction : instead of a text, or instead of many one hot text features, i had only a continuous feature with this score, usefull.\n\nThis ClaimDescription score was build separately for each fold **only on training sets during CV** in a pipeline of transformations, and before LGBM or catboost or NN trainings. \nThere would be target leakage if i had trained this ClaimDescription score on training sets **and on valid sets** before training LGBM, NN etc. in CV , but I didn't.\nAnd those transformers, transformed later valid samples (and test samples of course) just before predictions. Those valid and test samples were not used during training of transformers.\n\nAlso, if **CV RMSE** (RMSE on valid sets) is lower on my final LGBM when I'm using the ClaimDescriptionScore feature, than when I'm not, **I'm almost certain that this ClaimDescription Score will improve the LGBM RMSE on the PRIVATE leaderboard**. \n**Never mind if the RMSE is higher on PUBLIC leaderboard** ; anyway, I didn't submit this little LGBM difference on public LB.\n\nSince the beginning of the challenge, I've made all my decisions by this way (number of topics with LDA implementation etc.). That's why I had submitted so few. At the end of the challenge, all my new CV RMSE on new ideas told me that those new ideas shouldn't improve my RMSE on unseen private leaderboard, so I didn't submitted them.\nThe [notebook](https://www.kaggle.com/adaubas/why-not-to-blend-catboost-extratree-and-a-nn) published during january is like that.\n\nThanks to scikit-learn for transformer classes and pipeline, there are very powerfull tools to make decisions : a [link](https://scikit-learn.org/stable/modules/compose.html#pipeline-chaining-estimators). ", "votes": 2 }, { "id": 1274310, "postDate": "2021-04-15T07:20:53.147Z", "content": "Thank you Ald. 👍
", "rawMarkdown": "Thank you Ald. 👍 \n", "votes": 1 } ] }, { "id": 1275331, "postDate": "2021-04-16T07:26:55.110Z", "content": "amazingly done buddy!
", "rawMarkdown": "amazingly done buddy!" } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:14:48Z", "competition_id": 24764, "competition_slug": "actuarial-loss-estimation", "team_id": 6113538, "solution_url": "/competitions/actuarial-loss-estimation/writeups/aldparis-14th-place-solution" }