(1/6) 1 month ago, @mevalphaleak hosted a machine learning competition alphamev.ai/. The goal? "Build a model for predicting back-runnable [MEV] transactions and their value." It just wrapped up and we wrote about it! Some key points ­čžÁ blog.credmark.com/mev-mlcompÔÇŽ

5:45 PM ┬Ě Sep 13, 2021

2
6
0
16
(2/6) We placed 3rd! Public results don't filter with these controls, so on early leaderboards we were in 10th. That means that 7 submissions with higher scores were actually overfit, but not us! Another case of "best-fit in the past" being "inflexible for the future"
1
0
0
3
(3/6) We experimented with 3 types of models: Artificial Neural Networks, Support Vector Machines, and Tree Based Algorithms. ANN and SVM proved to underperform when put head to head with our Tree Based Algorithm, which has superior categorization, so we continued with that.
1
0
0
4
(4/6) The primary reason was that we wanted to see if we could prove out one of our core assumptions: If we can abstract away the contrivances of blockchain data, can we be inclusive of a much larger data community?
1
0
0
3
(5/6) Our data engineer had never heard of MEV, yet was able to place 3rd out of 50. As far as we're concerned, this proves our assumption. We now have empirical evidence that we can leverage a wider range of developers, modelers, and other technical people to run our platform.
1
0
0
4
(6/6) This gives us a great template for our platform moving forward. We'll be announcing our first hackathon very shortly! Join our Discord to stay informed: discord.gg/smtnSHZq
0
0
0
4