Short-Term Stock Modeling
A student project on short-term market signals, backtesting, and feature engineering
Project Framing
The screenshot above shows the live output of one of the paper-trading versions of the project generating buy and short ideas for the day. This started as a fairly open-ended question about whether short-horizon price movement could be modeled in a way that was still somewhat interpretable, and it ended up turning into a lot of data collection, feature engineering, backtesting, and trial and error with different model families.
Methodology & Architecture
- Model selection: I tested linear baselines, random forests, SVMs, CNN-style experiments, and boosted trees. CNNs and SVMs mostly overfit the data I had, so the more useful path ended up being a simpler feature-engineering workflow with tree-based models.
- Final direction: after using random forests to sanity-check features, I moved to XGBoost as the final model family because it improved performance without making the workflow too opaque.
- Tuning: I used repeated cross-validation plus parameter search to narrow down a workable configuration instead of trusting a single train/test split.
- Data workflow: a lot of the project was ordinary but important data work: collecting bars, computing indicators, storing larger training tables in parquet, and keeping CSV copies around while debugging.
Execution Layer
I also connected the project to the Alpaca API for paper-trading experiments. That part was mainly a way to test the mechanics of a live pipeline. It was not used for live trading with real funds.