The gist of why this is so fast is:

1. For the vast majority of queries, we only need to compare hashes of the time series, not the time series themselves.

2. We make no attempt to detect shifted correlations. The correlations you see are all synchronized in time.