Even vs uneven AR(1) estimation

[CommonClimate]

This is more of a scientific question than a coding issue, but I flag it here for future reference.

Right now ar1_fit() uses two modalities to estimate the persistence parameter necessary to run ar1_sim(), and all the applications that depend on it:

• if the data are perfectly even, it uses the statsmodels ARMA fit.
• if data are not evenly spaced, it uses TAUEST, which is heavily biased.

No doubt there are situations where TAUEST is the only way to go. However, the statsmodels ARMA fit module is tolerant to missing values, so in the case where the time axis is uniform and there aren't too many missing values in the series (e.g. layer-counted records with only a few gaps), sticking to the ARMA fit would be preferable to TAUEST.

How to automate the decision? That is an open question at present, but it would enhance Pyleoclim to have a better criterion.

[khider]

Distribution of increments (big gaps vs smaller gaps). Important for AR1 fit, maybe able to use statsmodel. Warning: is_evenly_spaced used for autoTS sending. Do not change this function. Write a new function for tolerance or pass 'tol' as a new parameter.