I keep draw histories to use for developing systems. I do not view these as a model of the data produced by the mechanical or PRNG process (or in PA, BOTH). I view it as a collection of the results they paid on.

Every single state lottery runs test draws, some even publish the results. I place no value in this data because the state never intended to pay on those results. Power Ball used to publish test draw data.

Up until the end of last year, I excluded PA double draw data (where they picked an extra draw on some nights in the pick N games for a promotion) but I keep it in now because they paid on it. It was quite the project re doing every single history file I had with the new web scraper app. I made 2 versions. The first built the base file, with the header rows from day one of the game history to current. The second simply appends new data to those csv files. This replaced hours of manual copying into the files, particularly if I had not played a game for a long time (like pick 5) and needed to add years of data. Now, there are zero human transcription errors in the data sets, and updates happen with the click of a button... 2 minutes later 14 game histories are updated accurately!

I also maintain separate files for bonus balls like the mega ball and power ball.

Access to the history in or out of the app is a click away, this was the vision realized! 

So when processing this data, I am mostly looking for any micro biases in each column that hopefully points to the next number prediction that will pay out.

There are some that hold zero stock in the usefulness of past draw data, and they make compelling arguments, but I see it differently... there might be a connection, and even if it is coincidental at best, it is better than nothing.

This Markov Chain follower back test script may end up to be the single most complex function I will write, and even if there is no yield, I have honed some skill at trying to solve impossible problems with code, skills that will eventually transcend the lottery hobby.