need to estimate model on eMid data

[domenicodigangi]

• [x] Set up mlflow experiment to compare runs
• [x] Set up single run script and check that sequential starting values are s.t. logl(prev mod final) == logl(curr mod zero)
• [x] Add pox to load starting point from less complex model
• [x] Sequentially estimate model's versions of increasing complexity (done but it probably does not makesense. not using it)
• [x] Compare BIC for each solution
• [x] Out of sample models' comparison
• [x] weighted version
• [x] estimate models with previous time mat as regressors
• [x] estimate models with interest rate and previous time mats as regressors

[domenicodigangi]

need to check:

• [x] BiC is equal to loglikelihood
• [x] incoherent starting values in sequential sd estimates
• [x] rolling of the sd before plotting to test set

[domenicodigangi]

Run an exploratory comparison of optimization algos - ADAM, ADAMHD, LBFGS - across different model's settings.

• LBFGS consistently underperforms the other two. Hence I will leave it aside for future estimates
• ADAMHD always requires less iterations than ADAM (time per iter should be approx the same)
• ADAMHD always reaches better likelihoods, although the two are always very close.

[domenicodigangi]

Models' comparison considering only eonia rate as regressor:

• Binary : BIC suggests model with one reg, out of sample model with no reg but they are all very close.
• Weighted: BIC and out of sample log mse suggest model with 2N regressors

Attached csv from mlflow with runs results runs (3).csv

[domenicodigangi]

The current state of the scripts for emid estimates is more sufficient for the moment