{
  "id": 494091,
  "title": "Question on using future external data",
  "url": "/competitions/home-credit-credit-risk-model-stability/discussion/494091",
  "author_name": "Cael Hasse",
  "post_date": "2024-04-16T00:18:56.331000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I have been considering using external data whereby for each date decision, the new features contain data from after the decision. Maybe I have missed something but I don't find anything in the rules that forbids this but depending on how the competition creators intend on using the winning models, such an approach may render the model useless: If you want to calculate a probability of default at the time of decision, then future data is unavailable, rendering the model useless but if you wanted to update probabilities after the time of decision, the model could potentially be useful.</p>\n<p>Any clarification would be appreciated.</p>",
  "messages": [
    {
      "id": 2754301,
      "postDate": "2024-04-16T00:18:56.330Z",
      "content": "<p>I have been considering using external data whereby for each date decision, the new features contain data from after the decision. Maybe I have missed something but I don't find anything in the rules that forbids this but depending on how the competition creators intend on using the winning models, such an approach may render the model useless: If you want to calculate a probability of default at the time of decision, then future data is unavailable, rendering the model useless but if you wanted to update probabilities after the time of decision, the model could potentially be useful.</p>\n<p>Any clarification would be appreciated.</p>",
      "rawMarkdown": "I have been considering using external data whereby for each date decision, the new features contain data from after the decision. Maybe I have missed something but I don't find anything in the rules that forbids this but depending on how the competition creators intend on using the winning models, such an approach may render the model useless: If you want to calculate a probability of default at the time of decision, then future data is unavailable, rendering the model useless but if you wanted to update probabilities after the time of decision, the model could potentially be useful.\n\nAny clarification would be appreciated.",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2754301": "I have been considering using external data whereby for each date decision, the new features contain data from after the decision. Maybe I have missed something but I don't find anything in the rules that forbids this but depending on how the competition creators intend on using the winning models, such an approach may render the model useless: If you want to calculate a probability of default at the time of decision, then future data is unavailable, rendering the model useless but if you wanted to update probabilities after the time of decision, the model could potentially be useful.\n\nAny clarification would be appreciated."
  }
}