![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F761268%2F1df25298829a2c39b521a9c1cc2b3def%2FScreenshot%202025-07-03%20at%202.35.27AM.png?generation=1751490410446025&alt=media\")
I think the top winning solutions from past year could be a reasonable starting point and baseline for this year's challenge, as the data has literally the same format.
\nHowever, the changes introduced by the host add significant new challenges:
Star spectra
\nIn 2024, all observations were simulations using just 2 reference stars, I don't remember if the test set had different stars but maybe just a couple more. This allowed to characterize pretty well the spectra of the star out-of-transit since you could combine lots of observations and use PCA or other methods.
\nHowever in 2025 data, every observation is a different star. We are given some physical parameters of the star (mass, radius, temperature) that can be use to narrow down the possible spectras based on physical priors.
Limb darkening
\nSimulating limb darkening means that the light curve during transit is not constant but concave since the stellar surface is not uniformly bright: the center is brighter than the edge (or \"limb\").
\nThere are different parametric models that are used to simulate limb darkening, so trying to fit them or reverse-engineer what has been used for generating the data should help.
Others
\nHost mention other changes like using more atmospheric profiles (from 3 to 12), upgrades in the physical models, and some undisclosed changes in the test set. I'm eager to see in the upcoming days what's the CV-LB correlation that people get.
I think the top winning solutions from past year could be a reasonable starting point and baseline for this year's challenge, as the data has literally the same format.
\nHowever, the changes introduced by the host add significant new challenges:
Star spectra
\nIn 2024, all observations were simulations using just 2 reference stars, I don't remember if the test set had different stars but maybe just a couple more. This allowed to characterize pretty well the spectra of the star out-of-transit since you could combine lots of observations and use PCA or other methods.
\nHowever in 2025 data, every observation is a different star. We are given some physical parameters of the star (mass, radius, temperature) that can be use to narrow down the possible spectras based on physical priors.
Limb darkening
\nSimulating limb darkening means that the light curve during transit is not constant but concave since the stellar surface is not uniformly bright: the center is brighter than the edge (or \"limb\").
\nThere are different parametric models that are used to simulate limb darkening, so trying to fit them or reverse-engineer what has been used for generating the data should help.
Others
\nHost mention other changes like using more atmospheric profiles (from 3 to 12), upgrades in the physical models, and some undisclosed changes in the test set. I'm eager to see in the upcoming days what's the CV-LB correlation that people get.
There also seem to be \"cropped\" transits, where transit starts before the exporsure. Like planet 2554492145
", "rawMarkdown": "There also seem to be \"cropped\" transits, where transit starts before the exporsure. Like planet 2554492145", "votes": 1, "replies": [ { "id": 3239475, "postDate": "2025-07-02T21:07:18.753Z", "content": "@sakvaua we seen similar planet in 2024 too right, didn't remember but seen similar cropped transit.
\n
Thanks for sharing! This makes the problem so hard.
", "rawMarkdown": "Thanks for sharing! This makes the problem so hard." }, { "id": 3239823, "postDate": "2025-07-03T07:27:55.270Z", "content": "And also \"spiky\" ones like 3649218579, 360841026, etc.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F349155%2F668a5ad76fc5e1c2128e7649a41e1304%2Fspiky.png?generation=1751527673055078&alt=media\")
\n\nHowever in 2025 data, every observation is a different star.
\n
At least, this makes CV a bit easier this time 😄
", "rawMarkdown": "> However in 2025 data, every observation is a different star.\n\nAt least, this makes CV a bit easier this time 😄" } ], "comments": [ { "id": 3239380, "author_name": "DennisSakva", "author_url": "", "post_date": "2025-07-02T19:06:21.460000", "content": "There also seem to be \"cropped\" transits, where transit starts before the exporsure. Like planet 2554492145
", "votes": 1, "replies": [ { "id": 3239475, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2025-07-02T21:07:18.753000", "content": "@sakvaua we seen similar planet in 2024 too right, didn't remember but seen similar cropped transit.
\n
Thanks for sharing! This makes the problem so hard.
", "votes": 0, "replies": [] }, { "id": 3239823, "author_name": "DennisSakva", "author_url": "", "post_date": "2025-07-03T07:27:55.270000", "content": "And also \"spiky\" ones like 3649218579, 360841026, etc.
\n
\n\nHowever in 2025 data, every observation is a different star.
\n
At least, this makes CV a bit easier this time 😄
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3238985": "I think the top winning solutions from past year could be a reasonable starting point and baseline for this year's challenge, as the data has literally the same format.\nHowever, the changes introduced by the host add significant new challenges:\n\n**Star spectra**\nIn 2024, all observations were simulations using just 2 reference stars, I don't remember if the test set had different stars but maybe just a couple more. This allowed to characterize pretty well the spectra of the star out-of-transit since you could combine lots of observations and use PCA or other methods.\nHowever in 2025 data, **every observation is a different star.** We are given some physical parameters of the star (mass, radius, temperature) that can be use to narrow down the possible spectras based on physical priors.\n\n**Limb darkening**\nSimulating limb darkening means that the light curve during transit is not constant but concave since the stellar surface is not uniformly bright: the center is brighter than the edge (or \"limb\").\nThere are different parametric models that are used to simulate limb darkening, so trying to fit them or reverse-engineer what has been used for generating the data should help.\n\n**Others**\nHost mention other changes like using more atmospheric profiles (from 3 to 12), upgrades in the physical models, and some undisclosed changes in the test set. I'm eager to see in the upcoming days what's the CV-LB correlation that people get.\n\n", "3239380": "There also seem to be \"cropped\" transits, where transit starts before the exporsure. Like planet 2554492145", "3243967": "> However in 2025 data, every observation is a different star.\n\nAt least, this makes CV a bit easier this time 😄" } }