Howdy!
This is the repository (just a notebook, really) tied to the winning submission for the TAMUDatathon 2020 "Stock Prediction" challenge. Further details, including the provenance of the challenge, can be found at the formal devpost.com submission.
The crux of the challenge is that we have daily mystery stock data, a bunch of covariates, and we want to predict next day's return. The twist is that for a consequent buy/sell, we need to place a "fraction" for how much to stake our capital.
For some more context, this was a competition held for undergraduate students, graduate students, and people who gradudated within a year from when the competition was held. As such, I technically participated while being a master's student in Statistics from Texas A&M.
This submission/solution ended up being the winning approach, but there are a litany of reasons why it may not be realistic, including but not limited to liquidity, what's the actual spread, no transaction costs were imposed, etc. Even more so, there's a lot of variance in this whole process that consequently gives you a fighting chance, even with a sub-optimal submission. That being said, imposing and/or correcting for all of "that" would potentially have made the challenge less fun for others and just more things to deal with in a 24 hour period. In any case, I had fun tooling around and messing with this.
Also, to get the most out of this notebook, you shouldn't really focus too much on the "stock" aspect and instead think of the following as a template for a supervised learning problem and an example of how one constructs pipelines and related items. Really, the way I treated this, was I'm given "X" and "y", now go build a good predictor and everything else was auxiliary.