"We’re also going to run into percent colinearity in the X variable if we run these scenarios with constant block rewards because the feature vectors will either be identical or we will have extremely close vectors with low variation.
If difficulty changes over epochs that should help in overall updating but we may run into issues on the individual epochs being so related that no good inference can be drawn. Do we have any way to tackle that issue? I could just toggle up variation in inter prime block difficulties to give some spread but I’m not sure if that will be enough if the spread isn’t rather wide"
"Right now, given the perfectly colinear X, we end up getting very negative beta values that are not at all near the population because it is essentially just biasing towards whichever side is >= 50% probability likelihood. But I'll see if it converges when randomness and block difficulty changes are implemented"
From comments in slack:
"We’re also going to run into percent colinearity in the X variable if we run these scenarios with constant block rewards because the feature vectors will either be identical or we will have extremely close vectors with low variation. If difficulty changes over epochs that should help in overall updating but we may run into issues on the individual epochs being so related that no good inference can be drawn. Do we have any way to tackle that issue? I could just toggle up variation in inter prime block difficulties to give some spread but I’m not sure if that will be enough if the spread isn’t rather wide"
"Right now, given the perfectly colinear X, we end up getting very negative beta values that are not at all near the population because it is essentially just biasing towards whichever side is >= 50% probability likelihood. But I'll see if it converges when randomness and block difficulty changes are implemented"