Actually, I use both but my production models are in Clojure.
I often end up implementing minor things myself using lower level abstractions (e.g., Linear Regressions or PCA with whitening using Matrix libraries) and I check the results and/or try new things using scikit-learn.
So in general, I'd say I do the programming (outputing intermediate CSVs, tests, web service, thread handling, UI, ...) in Clojure(Script), and try other approaches (e.g., other models/parameters/...) in Python.
I'm quite happy with this pipeline but probably to some extent because I really love to understand how things work and nothing pushes you to learn as much as a missing function in your ML library :-)
- a number of smart features (usually a few k) depending on the series (using lags, aggregates, curve fits, combinations of features, ...)
- an iterative algorithm that selects features using maximum relevance (~ correlation with the target) / minimum redundancy and adds them to the model
- simple pca and ridge regression (because it's fast)
- a few optimizations of the final model (removing features, selecting a better ridge regression alpha with CV, ...)
The stack is pure Clojure / Clojurescript.