This project explores how a Temporal Convolutional Network (TCN) can be used in a stock market setting.
The core focus is a universe of 50 stocks, with the model producing daily signals that can be reviewed over time.
The project is designed as a full research workflow: running experiments, comparing outcomes across runs, and tracking how behavior changes across different training settings.
It is intended as a practical showcase of applying machine learning to market data in a structured, repeatable way.
TCNs are well-suited for time-series problems because they use 1D convolutions across temporal data, which makes sequence modeling efficient.
Dilated convolutions let the model capture both short-term and longer-range patterns without needing very deep recurrent stacks.
Because these convolutions are causal in time-series setups, they align naturally with forecasting tasks where future information should not leak into current predictions.
In practice, this gives a good balance of speed, stability, and pattern-detection power for market signal modeling.
For public viewing, experiment runs and tracked metrics are available in MLflow at tcnshowcase.sites.tjhsst.edu.
To view data, you can click on Experiments on the left, then click on any listed experiment (or check multiple different experiments and click Compare), and switch to the chart section. Afterwards, you can expand the left side with "load more", change the eye icon setting to "Show all", and pick items from the Group By tab, which includes parameters such as projection tempature, data file path, etc. to see the differences in performance between different model configurations.
Users can also download mlflow.sql.xz and run restoredb.py to convert it into an sqlite3 database, which can then be viewed with MLflow on a local machine.