Machine Learning for Automated Trading

Background

Trading is an appealing domain for a student project: the data is plentiful, the objective is unambiguous, and the results are easy to see. It is also a domain where the naive approach fails instructively.

The aim is not to build a money-printing bot. It is to understand what actually limits learning in this setting — data leakage, regime change, overfitting to backtest noise, and the gap between a good backtest and a deployable system.

What you would build

You would work through the full pipeline rather than one stage:

• acquiring and cleaning market data into a time-series store,
• building and tuning models with proper experiment tracking,
• analysing anomalies and failure cases rather than just headline metrics,
• presenting results so that a human can actually interrogate them.

Visualisation is doing real work here, not decoration: the hard question is how to show a strategy’s behaviour over time in a way a person can reason about.

Background reading

Start by reading about look-ahead bias and overfitting in financial ML before building anything. Most of the interesting findings in this project come from encountering those problems directly.