← 1st Round Findings

Hypothesis

Coins are assumed to have different relevant features.

We use hierarchical clustering on concurrent log returns to find similarities and groups which would be assumed to share the same dynamics.

Outcome Paths

If groups not found:

  • Look for another way to cluster (e.g. GMM that clusters based on distributions)
  • Brute force groups through features selection per coin and group by that instead of indirect measures like return correlation
  • OR use generalized features proven beyond reasonable doubt to represent the dataset

If groups found:

  • Examine the feature groupings and sanity check them
  • Run a feature brute force to detect the most significant relations
  • See if the types of features and the correlation changes across groups

Result

Not enough time to sort coins into proper groups - that would require training multiple models on different groups which would need to be tested and documented with each model taking more time to optimize.

While a good idea that would make sense to pursue to improve performance, we didn’t have enough time for implementation. Other higher priority workflows took precedence.