eigs stabilizer
moonlians opened this issue · comments
My predictions are diverging from the data after the timeseries ends, so I tried the tuner.stabilize feature.
However, it seems like I can't change the eigenvalues a small enough amount - it places them close to 1, but it's still not optimised for prediction, it's over or under shooting. Is there a more fine tuned way of changing them?
I was also wondering, since the eigenvalues, modes, and dynamics change slightly based on the starting point in time, and some are better at predicting than others, whether there's some optimisation already written where I could just find the best of each and then use those? So, for instance, then enforcing dynamics and eigenvalues, but allowing modes to vary at different starting points, if that might help with stability.
This second thing isn't going to work, I think - I tried regressing the eigenvalues and quality of predictions, and there was no relationship, not even one that looked non-linear. So maybe there isn't a way to optimise them for prediction anyway, and I just stick with the short term predictions... they are anyway pretty good!