Nice job -- I have examples using statsmodels for similar (not time series) data [1,2]. I typically use this for EDA before regression modelling, so dependencies in that scenario are not a big deal. But I might weep if someone told me no scipy in production.