![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18102%2F1a011ee06339075039995790bd86b046%2FScreenshot%202020-05-31%20at%2010.32.51.png?generation=1591101401062776&alt=media\")
LGBMs with quantile regression were trained on time-series and geo features for short term predictions. Aggressive 1w average smoothing was used for long term predictions. Due to the large number of small locations the top 30 country/state had to be adjusted manually.
\n\nNotebooks\n\nhttps://www.kaggle.com/gaborfodor/c19w5-create-submission\nhttps://www.kaggle.com/gaborfodor/c19w5-check-submission\nhttps://www.kaggle.com/gaborfodor/c19w5-train-lgbs
\n\nRescaled and rounded features to 1-2 decimals to decrease overfitting
\n\nI started with searching public us county level demographic data (age, income, population density, covid19 lockdown info etc.) I found a few useful sources though I did not have time to clean and merge them. The only external data I used was the geo encoded lat-lon coordinates for each location.
\n\nFor each target/quantile/forecast lag separate model was trained \nwith location based 5-fold CV and early stopping based on pinball loss.\nModels were only trained to predict the next 1-14 days.
\n\nY[t] * 0.66 + Y[t-1] * 0.33)![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18102%2F1788c82f1c31a3e3900073134256424a%2FScreenshot%202020-06-02%20at%2014.39.01.png?generation=1591101619319599&alt=media\")
Good stuff. I believe you have a high chance of winning this!
", "votes": 3, "replies": [ { "id": 873523, "author_name": "beluga", "author_url": "", "post_date": "2020-06-04T08:14:53.357000", "content": "Thanks, I don't think I will be able to reach you, but I hope to get closer ;)
", "votes": 2, "replies": [] }, { "id": 873524, "author_name": "Psi", "author_url": "", "post_date": "2020-06-04T08:16:29.070000", "content": "Still enough time left to catch us, on good days you gain up to 0.001 against us and on bad days you lose little, and usually your model gets better towards end of week.
", "votes": 2, "replies": [] }, { "id": 873554, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2020-06-04T09:05:18.313000", "content": "My bet is:\n1) beluga\n2) DiDoKaKaKaPs\n3) disinfectant infection
", "votes": 4, "replies": [] }, { "id": 873556, "author_name": "Psi", "author_url": "", "post_date": "2020-06-04T09:06:06.897000", "content": "Same here - the battle might be actually for the 2nd place
", "votes": 1, "replies": [] }, { "id": 873883, "author_name": "beluga", "author_url": "", "post_date": "2020-06-04T13:50:33.073000", "content": "It will be close for sure\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18102%2F9d4c5fb091a0ea296af174b8601d6efc%2FScreenshot%202020-06-04%20at%2015.48.39.png?generation=1591278553258974&alt=media\")
Maybe\n1) beluga\n2) disinfectant infection\n3) DiDoKaKaKaPs\n4) ML Alcohol Sanitizer
", "votes": 2, "replies": [] }, { "id": 874168, "author_name": "Psi", "author_url": "", "post_date": "2020-06-04T17:20:08.030000", "content": "@beluga I didn't study the different models as well as you did, are you aware of which countries you are doing so much better currently than others?
", "votes": 1, "replies": [] }, { "id": 874209, "author_name": "beluga", "author_url": "", "post_date": "2020-06-04T17:52:52.457000", "content": "Not clear. I have the worst score for Brazil among the top submissions, excluding Brazil I would be first, focusing only on US 3rd. I have good guesses for Russia, Mexico. https://www.kaggle.com/gaborfodor/w5-top-submissions has a section for the unexpected regions for each team.
", "votes": 0, "replies": [] }, { "id": 874215, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2020-06-04T17:57:33.897000", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F471945%2Fe91ed5db817bf61299e81a94cba0524f%2FCapture.PNG?generation=1591293358837631&alt=media\")
Lockdown / Northquay | DiDoKaKaKaPs | beluga | disinfectant injection
\n\nI have sorted the losses by std. These are the top 20 that make the difference today (reverse order)
", "votes": 3, "replies": [] }, { "id": 876215, "author_name": "Psi", "author_url": "", "post_date": "2020-06-06T14:37:36.187000", "content": "There we go :)
\n\nWill be close between 2-4 now I guess, anything can happen.
\n\nRegardless of the results here, it is really devastating what is going on in Brazil and other countries. Worldwide maximum number of new cases yesterday again.
", "votes": 0, "replies": [] }, { "id": 876227, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2020-06-06T14:43:35.080000", "content": "Brazil data came zero for some reason. I guess they entered it late.
", "votes": 0, "replies": [] }, { "id": 876368, "author_name": "Kirderf", "author_url": "", "post_date": "2020-06-06T16:37:21.023000", "content": "@philippsinger It really is a tragedy what happens, going in the wrong direction in many countries, our thoughts go to them, now we hope that quickly gets better.
", "votes": 1, "replies": [] }, { "id": 876394, "author_name": "Psi", "author_url": "", "post_date": "2020-06-06T16:50:32.043000", "content": "@aerdem4 Really? I saw it way over 15k before midnight, so that should not happen...
\n\nHas a huge impact on this competition of course :/
", "votes": 0, "replies": [] }, { "id": 876415, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2020-06-06T17:06:43.627000", "content": "This competition has large noise based on when countries report and data is updated. Also some countries are updating their counts retroactively but since Kaggle calculates everything as difference on cumulative sum, we get huge spikes time to time.
", "votes": 0, "replies": [] }, { "id": 876419, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2020-06-06T17:08:32.017000", "content": "So tomorrow's update will be 2x Brazil probably. For my model, 2 1x Brazil is better than 1 0x and 1 2x.
", "votes": 0, "replies": [] }, { "id": 876442, "author_name": "Psi", "author_url": "", "post_date": "2020-06-06T17:26:05.030000", "content": "Actually what might happen is that JHU fixes the issue, and only the correct ones from today will be added, and the other backported to yday. These lost counts wont find their way to kaggle scoring though, as only extra day is always added.
", "votes": 0, "replies": [] }, { "id": 877337, "author_name": "Psi", "author_url": "", "post_date": "2020-06-07T14:15:12.590000", "content": "Regarding Brazil:\nhttps://www.aljazeera.com/news/2020/06/brazil-stops-publishing-coronavirus-deaths-infections-200607064204479.html
", "votes": 0, "replies": [] } ] }, { "id": 875131, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2020-06-05T14:44:03.280000", "content": "I think having no model after 14 days and using the average of previous predictions was a good idea. While training my NN, I had noticed that it gets easier to overfit after 14 days and solved the issue by just training less epochs. But yours seems to be a better solution.
", "votes": 4, "replies": [ { "id": 875232, "author_name": "Psi", "author_url": "", "post_date": "2020-06-05T15:34:15.310000", "content": "Yeah, his model seems to be doing specifically better after those 14 days than others.
", "votes": 1, "replies": [] }, { "id": 875294, "author_name": "Kirderf", "author_url": "", "post_date": "2020-06-05T16:24:52.160000", "content": "Smart idea, with split prediction.
", "votes": 0, "replies": [] } ] } ], "comment_count": 21, "messages": [ { "id": 871544, "postDate": "2020-06-02T12:54:10.903Z", "content": "LGBMs with quantile regression were trained on time-series and geo features for short term predictions. Aggressive 1w average smoothing was used for long term predictions. Due to the large number of small locations the top 30 country/state had to be adjusted manually.
\n\nNotebooks\n\nhttps://www.kaggle.com/gaborfodor/c19w5-create-submission\nhttps://www.kaggle.com/gaborfodor/c19w5-check-submission\nhttps://www.kaggle.com/gaborfodor/c19w5-train-lgbs
\n\nRescaled and rounded features to 1-2 decimals to decrease overfitting
\n\nI started with searching public us county level demographic data (age, income, population density, covid19 lockdown info etc.) I found a few useful sources though I did not have time to clean and merge them. The only external data I used was the geo encoded lat-lon coordinates for each location.
\n\nFor each target/quantile/forecast lag separate model was trained \nwith location based 5-fold CV and early stopping based on pinball loss.\nModels were only trained to predict the next 1-14 days.
\n\nY[t] * 0.66 + Y[t-1] * 0.33)
Good stuff. I believe you have a high chance of winning this!
", "rawMarkdown": "Good stuff. I believe you have a high chance of winning this!", "votes": 3, "replies": [ { "id": 873523, "postDate": "2020-06-04T08:14:53.357Z", "content": "Thanks, I don't think I will be able to reach you, but I hope to get closer ;)
", "rawMarkdown": "Thanks, I don't think I will be able to reach you, but I hope to get closer ;)", "votes": 2 }, { "id": 873524, "postDate": "2020-06-04T08:16:29.070Z", "content": "Still enough time left to catch us, on good days you gain up to 0.001 against us and on bad days you lose little, and usually your model gets better towards end of week.
", "rawMarkdown": "Still enough time left to catch us, on good days you gain up to 0.001 against us and on bad days you lose little, and usually your model gets better towards end of week.", "votes": 2 }, { "id": 873554, "postDate": "2020-06-04T09:05:18.313Z", "content": "My bet is:\n1) beluga\n2) DiDoKaKaKaPs\n3) disinfectant infection
", "rawMarkdown": "My bet is:\n1) beluga\n2) DiDoKaKaKaPs\n3) disinfectant infection", "votes": 4 }, { "id": 873556, "postDate": "2020-06-04T09:06:06.897Z", "content": "Same here - the battle might be actually for the 2nd place
", "rawMarkdown": "Same here - the battle might be actually for the 2nd place", "votes": 1 }, { "id": 873883, "postDate": "2020-06-04T13:50:33.073Z", "content": "It will be close for sure\n
Maybe\n1) beluga\n2) disinfectant infection\n3) DiDoKaKaKaPs\n4) ML Alcohol Sanitizer
", "rawMarkdown": "Maybe\n1) beluga\n2) disinfectant infection\n3) DiDoKaKaKaPs\n4) ML Alcohol Sanitizer\n", "votes": 2 }, { "id": 874168, "postDate": "2020-06-04T17:20:08.030Z", "content": "@beluga I didn't study the different models as well as you did, are you aware of which countries you are doing so much better currently than others?
", "rawMarkdown": "@beluga I didn't study the different models as well as you did, are you aware of which countries you are doing so much better currently than others?", "votes": 1 }, { "id": 874209, "postDate": "2020-06-04T17:52:52.457Z", "content": "Not clear. I have the worst score for Brazil among the top submissions, excluding Brazil I would be first, focusing only on US 3rd. I have good guesses for Russia, Mexico. https://www.kaggle.com/gaborfodor/w5-top-submissions has a section for the unexpected regions for each team.
", "rawMarkdown": "Not clear. I have the worst score for Brazil among the top submissions, excluding Brazil I would be first, focusing only on US 3rd. I have good guesses for Russia, Mexico. https://www.kaggle.com/gaborfodor/w5-top-submissions has a section for the unexpected regions for each team." }, { "id": 874215, "postDate": "2020-06-04T17:57:33.897Z", "content": "
Lockdown / Northquay | DiDoKaKaKaPs | beluga | disinfectant injection
\n\nI have sorted the losses by std. These are the top 20 that make the difference today (reverse order)
", "rawMarkdown": "\n\nLockdown / Northquay | DiDoKaKaKaPs | beluga | disinfectant injection\n\nI have sorted the losses by std. These are the top 20 that make the difference today (reverse order)\n", "votes": 3 }, { "id": 876215, "postDate": "2020-06-06T14:37:36.187Z", "content": "There we go :)
\n\nWill be close between 2-4 now I guess, anything can happen.
\n\nRegardless of the results here, it is really devastating what is going on in Brazil and other countries. Worldwide maximum number of new cases yesterday again.
", "rawMarkdown": "There we go :)\n\nWill be close between 2-4 now I guess, anything can happen.\n\nRegardless of the results here, it is really devastating what is going on in Brazil and other countries. Worldwide maximum number of new cases yesterday again." }, { "id": 876227, "postDate": "2020-06-06T14:43:35.080Z", "content": "Brazil data came zero for some reason. I guess they entered it late.
", "rawMarkdown": "Brazil data came zero for some reason. I guess they entered it late." }, { "id": 876368, "postDate": "2020-06-06T16:37:21.023Z", "content": "@philippsinger It really is a tragedy what happens, going in the wrong direction in many countries, our thoughts go to them, now we hope that quickly gets better.
", "rawMarkdown": "@philippsinger It really is a tragedy what happens, going in the wrong direction in many countries, our thoughts go to them, now we hope that quickly gets better.", "votes": 1 }, { "id": 876394, "postDate": "2020-06-06T16:50:32.043Z", "content": "@aerdem4 Really? I saw it way over 15k before midnight, so that should not happen...
\n\nHas a huge impact on this competition of course :/
", "rawMarkdown": "@aerdem4 Really? I saw it way over 15k before midnight, so that should not happen...\n\nHas a huge impact on this competition of course :/" }, { "id": 876415, "postDate": "2020-06-06T17:06:43.627Z", "content": "This competition has large noise based on when countries report and data is updated. Also some countries are updating their counts retroactively but since Kaggle calculates everything as difference on cumulative sum, we get huge spikes time to time.
", "rawMarkdown": "This competition has large noise based on when countries report and data is updated. Also some countries are updating their counts retroactively but since Kaggle calculates everything as difference on cumulative sum, we get huge spikes time to time." }, { "id": 876419, "postDate": "2020-06-06T17:08:32.017Z", "content": "So tomorrow's update will be 2x Brazil probably. For my model, 2 1x Brazil is better than 1 0x and 1 2x.
", "rawMarkdown": "So tomorrow's update will be 2x Brazil probably. For my model, 2 1x Brazil is better than 1 0x and 1 2x." }, { "id": 876442, "postDate": "2020-06-06T17:26:05.030Z", "content": "Actually what might happen is that JHU fixes the issue, and only the correct ones from today will be added, and the other backported to yday. These lost counts wont find their way to kaggle scoring though, as only extra day is always added.
", "rawMarkdown": "Actually what might happen is that JHU fixes the issue, and only the correct ones from today will be added, and the other backported to yday. These lost counts wont find their way to kaggle scoring though, as only extra day is always added." }, { "id": 877337, "postDate": "2020-06-07T14:15:12.590Z", "content": "Regarding Brazil:\nhttps://www.aljazeera.com/news/2020/06/brazil-stops-publishing-coronavirus-deaths-infections-200607064204479.html
", "rawMarkdown": "Regarding Brazil:\nhttps://www.aljazeera.com/news/2020/06/brazil-stops-publishing-coronavirus-deaths-infections-200607064204479.html" } ] }, { "id": 875131, "postDate": "2020-06-05T14:44:03.280Z", "content": "I think having no model after 14 days and using the average of previous predictions was a good idea. While training my NN, I had noticed that it gets easier to overfit after 14 days and solved the issue by just training less epochs. But yours seems to be a better solution.
", "rawMarkdown": "I think having no model after 14 days and using the average of previous predictions was a good idea. While training my NN, I had noticed that it gets easier to overfit after 14 days and solved the issue by just training less epochs. But yours seems to be a better solution.", "votes": 4, "replies": [ { "id": 875232, "postDate": "2020-06-05T15:34:15.310Z", "content": "Yeah, his model seems to be doing specifically better after those 14 days than others.
", "rawMarkdown": "Yeah, his model seems to be doing specifically better after those 14 days than others.", "votes": 1 }, { "id": 875294, "postDate": "2020-06-05T16:24:52.160Z", "content": "Smart idea, with split prediction.
", "rawMarkdown": "Smart idea, with split prediction." } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T12:19:54Z", "competition_id": 20475, "competition_slug": "covid19-global-forecasting-week-5", "team_id": 4866570, "solution_url": "/competitions/covid19-global-forecasting-week-5/writeups/beluga-1st-place-solution-lgbm-with-some-adjustmen" }