v3.0.3

Improved load balancing for multi-threaded code. For the symmetric case (corkendall(x)) this improved performance by as much as 25%.

v3.0.2

Changed multi-threading code in light of https://julialang.org/blog/2023/07/PSA-dont-use-threadid/

v3.0.0

• Simpler code, with the result that corkendall works for vectors of any type for which isless is defined. Hence:

julia> corkendall(["a","b","c"],["z","y","x"])
-1.0

• Changed corkendall's behaviour when skipmissing = :none to follow that of StatsBase.pairwise
• More tests.

v2.2.0

This version has:

• Improved test coverage.
• Passes tests on 32 bit OS's.
• Slightly more concise code.

Improved performance for small correlation matrix

This release improves performance when Threads.nthreads() is large but the correlation matrix being calculated has few rows and columns. The code no longer allocates scratch space that will never be used.

First threaded version to be considered for StatsBase

Release to allow for code review prior to making a PR to replace StatsBase's corkendall with the version from this package.

Advantages:

1. Speed. As described in the README file, this version of corkendall is faster than that in StatsBase by a factor close to the number of cores available. There is a speed up only when at least one of x and y is a matrix - threading is not (easily) possible in the vector-vector case.
2. The function has a new keyword argument skipmissing, with allowed values :none, :pairwise and :listwise

Drawback:

1. The function does not make use of StatsBase's pairwise function to implement the handling of missing values.

Version for 3rd PR to StatsBase

• Fixes for overflow errors, which were easy to hit on 32bit and possible to hit on 64bit
• Fixes for type instability.
• Improved tests.

Version for 1st PR to StatsBase

Fixed some type instabilities. This version is ready to be incorporated in StatsBase via a pull request.

See discussion at JuliaStats/StatsBase.jl#634

This PR was cancelled after tests revealed "overflow errors" on 32 bit

Faster again. Now uses less memory than StatsBase version.

1. Rewrote insertionsort! and mergesort! by copy-paste-edit from the relevant methods of sort! in base/sort.jl. Upshot was further speed improvements and reduction in memory footprint. The function now allocates 30% to 40% less memory than the StatsBase version.

   In the case of insertionsort! it might be possible to use sort! instead, but with a custom !isless(x,y) and that would avoid the unpleasant code duplication. But in the case of mergesort! I don't think that is possible as the function !isless would not "know" by how much to increase the swaps counter.

   Note that StatsBase.corkendall can also be criticised for for code duplication. It contains the mergesort algorithm in disguise as the function swaps!.

2. Moved function corkendallnaive to the tests folder.

A bit faster and uses less memory than previous release.

Minor speed improvement, allocates less memory that v1.1, but still more than StatsBase.corkendall.