Hi all
It’s now almost been a week since our launch and we’re very happy to see such vigorous engagement with the competition!
We’re very excited for our first weekly profit leaderboard on Saturday at 10pm CET!
I have 4 updates for you today:
- Choosing the submission you would like to use for the profit leaderboard (form available here)
- Added profitability feedback to the RMSE leaderboard
- Updated the RMSE-based evaluator
- Improved the starter-kit
1 Choosing the submission you would like to use for the profit leaderboard
You can now pick any of your previous submissions that you would like to enter for the Profit leaderboard. If you don’t make a specific choice, we will use your most recent successful submission for the profit leaderboard.
Please fill up this form to indicate your chosen submission id.
If you are part of a team, please nominate any one person from your team to fill it up.
The form closes at 10 PM CET on Saturday, 26th Dec
2 Profitability feedback on RMSE leaderboard
Upon every submission, we now include feedback regarding whether your model satisfies the non-negative training profit rule to make sure you are aware which of you submissions are eligible for the profit leaderboard. Note that models that do not have non-negative profit will not be included in the profit leaderboard.
For example, the mean model baseline (the template model) does not satisfy this rule and if I submit it, this is the warning I get (see the blue box):
3 Updated RMSE-based evaluator
A very helpful member of the community highlighted a particular exploit in our RMSE leaderboard which allowed the expected claim estimates predicted for past contracts, to be informed by their future.
Therefore, we are altering the RMSE leaderboard evaluator so that predictions for each year are made with access only to data from previous years. The RMSE is then computed by gathering those predictions and using the standard RMSE formula (for details see the overview page).
This way we ensure that, in the market-simulation, your models are able to use past data when available to evaluate the future, without allowing the future to predict the past.
4 Improved starter-kit
The starter-kit now includes:
- One new baseline model both as a
zip
and as Colaboratory notebooks(for both R and Python). - New
test.sh
andtest.bat
files that check yourzip
submissions before sending to AIcrowd (these might take a while to update) - New package loading functionality in Colaboratory for R users.
Good luck!
Ali