Yes, I'de expect the diagonals to be larger and it's odd in general that the outputs are so small. Below is the returns I gathered from NSR with my filters for each grade. Hopefully all I need more data, either quarterly or monthly. There is also a possibility that the small variance of returns by grade is resulting in this occuring, unlike a portfolio of stock, bonds, funds, ETFs, etc where returns fluctuate.

year A B C D E F G

2012 7.40% 12.48% 14.88% 17.60% 19.60% 21.22% 22.95%

2013 7.71% 12.12% 15.28% 18.04% 20.71% 22.49% 23.53%

2014 7.29% 11.37% 13.85% 16.63% 19.01% 22.43% 24.24%

2015 6.87% 10.33% 12.87% 15.74% 17.72% 20.47% 21.17%

st dev 0.30% 0.82% 0.94% 0.89% 1.08% 0.85% 1.14%

I didn't say they "would help". I'm not even an advocate of this particular approach. I was trying to answer your original question of whether someone had done this efficient frontier vs credit grades thing.

One of the blogs gives an example covariance matrix, which doesn't look like yours, so that's one piece of your puzzle. Of course he's doing a complex calculation to come up with these covariances, and you're likely taking a different approach.

In your matrix ... Wouldn't you expect the elements on diagonal be larger than the other elements on their row or column?