We are pleased to be able to announce the winner of the RecoGym challenge!

The winning team is "bx", which achieved an average click through rate of 2.19%! Congratulations!

Also congratulations to the four teams that tied for the 2nd place:

"Team MontiPuleri", "FabiCim", "HanShotFirst" and "AdremDataLab", all of which achieved a click through rate of 2.15%!

Now, for the prizes:

The 1000 Euro prize will be awarded to "bx" for the 1st place.

>Additionally, a 500 Euro prize will be awarded to each of the four 2nd place teams: "Team MontiPuleri", "FabiCim", "HanShotFirst" and "AdremDataLab".

Note: The difference between first and second place was so small that we required more than 100 simulations in order to accurately determine the winner. We will contact shortly the winning teams to discuss the payment details!

We would also like to thank all participants for all of the hard work and the enthusiasm you have shown in our first edition, we hope you have enjoyed the challenge! Once again, we apologise for the delays in announcing the final winners, the final evaluation took a lot more time than expected since the race was so tight! We would like to invite all participants to donate their entries so that others can learn for your submissions. Also, stay posted, we are preparing the 2nd edition of the challenge, with harder settings and bigger prizes!

In the meantime, for a more detailed analysis of the results:

In order to determine the winner we ran more than 100 simulations on the most competitive entries. A complete plot of each individual simulation is given below (the error bar represents a 95% interval each colour represents a different seed). Due to the large number of simulations it is difficult to see a clear trend here. Although it is clear that there is significant variation between seeds and the error bars of the individual seeds themselves are small.

Noting that the between seed error dominates we compute a standard error by computing the sample standard deviation and dividing by the square root of the sample size. We then plot 1 sigma error bars giving the following results:

We concentrated our simulations initially on the top 10 entries and gradually concentrated on the top entries, running more than 100 simulations on the top 6, which is why the error bars are much tighter on the right hand side. Zooming in on the highest performing entries reveals the following plot:

This plot reveals "bx" to have won the competition, but it is impossible to separate second place and so the judges have decided to award a 2nd place tie to "Team MontiPuleri", "FabiCim" and "HanShotFirst" and "AdremDataLab". Congratulations to everyone! See you next time!