I actually did something similar but less formal for an automated-composer experiment. I took several dozen tunes I liked, and extracted out the most common cho

I actually did something similar but less formal for an automated-composer experiment. I took several dozen tunes I liked, and extracted out the most common chord progressions found in them. I normalized it such that "key" was irrelalent because transposed keys are generally considered the same such that I was more concerned about the relative relationship between chords per song. After the analysis gave me my Markov-chain-like progression list, I randomly generated a some test tunes. While most of the results were agreeable, some I didn't like, and ended up trimming the offending chords out of my list. I realized I probably would need to analyze 3 levels of progressions -- a moving window of A-to-B-to-C -- to get better results, or at least better tests. I never finished that step. But I should point out that I was interested in generating progressions that sounded good to me. Ultimately I want to make tunes that I like even if I use a computer to help. That means I can use manual adjustments and/or backtracking the chain generation when needed. Therefore, going 3 levels was probably overkill. Leaving it at 2 gives me "happy accidents" anyhow.