{
  "fetched_at": "2026-10-19T15:37:06Z",
  "source_url": "https://www.kaggle.com/competitions/berkeleytradingcompetition2022/data",
  "pages": [
    {
      "name": "data-description",
      "content": "In train.csv, we provided information about each of the trades from our strategy. The features included are\n\n- pred edge: The predicted edge that our strategy generated. Predicted edge is in bips, where a bip is 1/100 of a percent. For example, if realized edge was 20 for a trade of $100 USD, our PNL would be $100*(20/10000) = $0.2\n- stock: Stock that was being traded (A, B, C, or D) \n- location: Exchange the trade took place (NYSE, NASDAQ, or CSE) \n- price: Price per share of the stock\n- size: Size of the transaction (in units of stock)\n- dir: Whether our strategy bought or sold (B or S)\n- counterparty: Who we traded against (Hedge Fund, Mutual Fund, or Retail) \n- sentiment: A score from -100 to 100 on how bearish/bullish retail investors are. You can think of this as sentiment data scraped from websites that retail traders frequent (Twitter/Reddit/etc)\n\nWe are trying to predict\n\n- realized edge: the realized edge of our strategy. Realized edge is in bips. \n- PNL: price * size * (realized edge / 10000) \n\nWhile developing the strategy, Traders hired some unpaid interns to scrape data on the counterparty and sentiment. However, because we never paid them, the interns did a bad job on recording the data and we believe that 30-40% of the data in those columns are incorrect. There will be separate auctions held for higher quality data on the counterparty and sentiment. \n\n\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Description",
      "content": "Traders at Berkeley recently created their own high frequency trading strategy, acting on 4 stocks traded across 3 markets. However, we have noticed that there has been a large discrepancy between our predicted and realized edge and want to understand where this slippage is coming from. Your goal is to predict the realized edge, or how much Traders is actually earning per trade, and use that information to decide which trades in the test set will have positive realized edge. We have provided relevant information about each trade, explained in the Dataset section. You may use any of the features provided.\n\nPlease see [GAME YOSEMITE CLARIFICATIONS](https://docs.google.com/document/d/1B7wEeQ8J0DPXunxw1tIRmUvhrST3LOXcG6bU2YoIK4g/edit?usp=sharing) for additional information on SUBMISSIONS, SCORING, and AUCTION.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Evaluation",
      "content": "The evaluation metric for this competition is [RMSE](https://en.wikipedia.org/wiki/Root-mean-square_deviation). \n\n## Submission Format\n\n**For every trade in the dataset**, submission files should contain two columns: `trade id` and `pred edge`. \n\n\nThe file should contain a header and have the following format:\n```\ntrade id,pred edge\n0,0.01563\n1,-0.0329\n```",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "rules",
      "content": "",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    }
  ]
}