ML Stock Signal Model

What it is

A machine learning model that produces a signal from historical market data, trained as a class project using course-provided resources, then deployed to AWS endpoints and connected to a small hosted app so it could actually serve predictions rather than sit in a notebook.

To be clear about what this is: the modeling work leaned on course materials and guidance. The part I drove was everything after training.

Why I built it

A model that only runs in a notebook has skipped most of the hard parts. I wanted to take something through to a real endpoint and find out what breaks in between — which turned out to be most of it.

What I learned

• Deployment is where the assumptions surface. Features that were trivial to compute over a static training set had to be computed on demand, from data arriving in a different shape. Reconciling those two was the bulk of the work.
• Endpoints cost money while they idle. Which is an obvious thing to say and a less obvious thing to plan for. It changed how I thought about the architecture.
• Backtest performance is not a promise. The model looks better on historical data than it has any right to look going forward, and understanding why that's true was probably the most valuable thing I took from it.

Where it stands

Complete as a class project. It's here because of what the deployment taught me, not as a claim that the model predicts anything useful.