{
  "id": 598156,
  "title": "Is the secret sauce intraday gradient updates? ",
  "url": "/competitions/jane-street-real-time-market-data-forecasting/discussion/598156",
  "author_name": "",
  "post_date": "2025-08-09T04:16:19.409155400Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Since we haven't seen other top solutions shared besides the 8th place solution, I am just speculating on what additional tricks are used to achieve a higher score. I think one thing we can do is the following: train a model to predict responder_6 at t-20. This model will have very high accuracy and give us pseudo-targets we can use to perform intraday gradient updates without waiting for the lagged labels the subsequent day.</p>",
  "messages": [
    {
      "id": "3266434",
      "postDate": "08/09/2025 04:16:19",
      "content": "<p>Since we haven't seen other top solutions shared besides the 8th place solution, I am just speculating on what additional tricks are used to achieve a higher score. I think one thing we can do is the following: train a model to predict responder_6 at t-20. This model will have very high accuracy and give us pseudo-targets we can use to perform intraday gradient updates without waiting for the lagged labels the subsequent day.</p>",
      "rawMarkdown": "Since we haven't seen other top solutions shared besides the 8th place solution, I am just speculating on what additional tricks are used to achieve a higher score. I think one thing we can do is the following: train a model to predict responder_6 at t-20. This model will have very high accuracy and give us pseudo-targets we can use to perform intraday gradient updates without waiting for the lagged labels the subsequent day.",
      "votes": null
    },
    {
      "id": "3266881",
      "postDate": "08/09/2025 21:07:08",
      "content": "<p>I've tried that with little success. Although i do think the secret sauce is good online learning, my solution was a simple MLP that didn't require offline training</p>",
      "rawMarkdown": "I've tried that with little success. Although i do think the secret sauce is good online learning, my solution was a simple MLP that didn't require offline training",
      "votes": null
    },
    {
      "id": "3266882",
      "postDate": "08/09/2025 21:17:13",
      "content": "<p>Do you train a new model every few days? I think even with a sequence model that is doable depending on the optimal number of epochs. </p>",
      "rawMarkdown": "Do you train a new model every few days? I think even with a sequence model that is doable depending on the optimal number of epochs.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3266881,
      "author_name": "johnpayne0",
      "author_url": "",
      "post_date": "08/09/2025 21:07:08",
      "content": "<p>I've tried that with little success. Although i do think the secret sauce is good online learning, my solution was a simple MLP that didn't require offline training</p>",
      "votes": null,
      "replies": [
        {
          "id": 3266882,
          "author_name": "xiaoz259",
          "author_url": "",
          "post_date": "08/09/2025 21:17:13",
          "content": "<p>Do you train a new model every few days? I think even with a sequence model that is doable depending on the optimal number of epochs. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3266434": "Since we haven't seen other top solutions shared besides the 8th place solution, I am just speculating on what additional tricks are used to achieve a higher score. I think one thing we can do is the following: train a model to predict responder_6 at t-20. This model will have very high accuracy and give us pseudo-targets we can use to perform intraday gradient updates without waiting for the lagged labels the subsequent day.",
    "3266881": "I've tried that with little success. Although i do think the secret sauce is good online learning, my solution was a simple MLP that didn't require offline training",
    "3266882": "Do you train a new model every few days? I think even with a sequence model that is doable depending on the optimal number of epochs."
  },
  "source": "meta"
}