While the Markov Chain approach seems much like follower data, the approach is different because it is capturing the observed frequencies and normalizing the distribution percentages.

This allows capturing a single pick, which can be done in a moving window, allowing for back testing the entire history of each column of each game. From this concept, pick data can be collected and compared to actual results, and an error function can be developed by comparing the pick to the next draw.

It is on this level that the agent which will initially parse the dictionary and generate a simple pick can be transformed into a more robust construct that can pick not only on the Markov Chain observed transition data, but to also "learn" by reinforcement (reward/penalty) how to adjust that pick to take into account the observed errors along the way.

The output will still be the same, one pick.

Over the years I have tried many things without fully understanding what I was looking for. I think the firm idea is not just static analysis, but also adding temporal context. I have hope that this type of analysis can help understand the "churn" better, or how patterns can repeat, just with different numbers in a given time frame.

I think my problem may have been trying to black box the parts of the problem I did not fully understand.

Version 1 will be with the simple dictionary parsing "agent", but that allows development of the smarter agent and the back test data for later refinement. It changes the process by only coding what I know until I understand the rest.

Pretty sure I can crank out an operating script for a basic Markov Chain pick in time to add it to the GUI app... now there will be three choices per game...

1. Classification

2. Followers

3. Markov Chain Pick

Both 2 and 3 will use the full history, classification will still be based on a sample.

That should provide plenty of coding work for the hobby in 2026.