Hey Dan!

First, sorry for the late reply. Your comment is a great one and I wanted to take the time to properly answer you and hopefully shed some light.

Also, thank you for such a contribution. I love having readers with your knowledge and discussing DS stuff.

Now, on to your questions.

1. Convolutional layer. As you know, a convolutional layer is mainly used to extract features from the data. It was perfect to deal with such a large number of player features that were being inputted, to form a more complex aggregation pattern compared to the team-level variables (which were simply sums or averages).

While it might not make sense to use it on non-image data at first, we're essentially doing the same here compared to what we would do if the input represented an image.

2. Regarding player injuries and trades. The order used to choose and place those 8 players is based on the average minutes per game they played in the previous 20 games (or so). I didn't mention it in the post to avoid making it too extensive, but I also worried about injuries and trades... That's why, at prediction time, the system I built took care of it by looking at the official injury reports (a PDF provided by the NBA) and also checking the official team rosters to discard those who weren't playing.

Take into account that the order changed from game to game, as it used rolling averages on minutes played as the sorting condition.

3. Point differential vs binary predictor. You're totally right here and, if I had to go through the process again (or resume it) I would probably follow the point differential approach. The reasoning behind using the sigmoid function in the output layer was to get a "confidence index", but using the point differential is definitely a better approach (saying it without having tested it yet).

However, I wasn't ignoring that info because point differential is dependent on points received and points scored, which were used as inputs.

4. An accuracy of 71.67% isn't as high. I mean, it's a great number but there are papers with similar numbers out there (I would have to look for them, but I assure you they exist). The key here is that accuracy is worthless when the goal is to make money betting, as your balance will depend on the correlation between your predictions and the bookies'. I hope, however, that point 2 answered your questions about how I managed the player-order doubts.

Hope I clarified your doubts. If not, please feel free to keep up the conversation either here or through LinkedIn! I think we could have a really interesting conversation :)

Let me know if you're still skeptical about these, I can promise it's all legit.

Thanks!