ML4T: Machine Learning for Trading

Georgia Tech OMSCS ML4T (CS 7646) connects machine learning with financial markets. The course is less about predicting prices and more about building decision pipelines that are statistically sound, risk-aware, and testable. This post summarizes what I learned in Fall 2024.

Course Focus

ML4T treats trading as a data and decision problem. You learn how to reason about signals, noise, and risk, then evaluate strategies under realistic constraints such as transaction costs and non-stationary markets.

Portfolio Theory: The Baseline

Modern Portfolio Theory introduces the risk-return tradeoff and the efficient frontier. The main lesson is that return alone is meaningless without risk, and diversification is not optional.

Key building blocks:

• Expected return and volatility
• Covariance and correlation
• Constraints that reflect real portfolios

Signals and Features

Raw prices are rarely enough. The course emphasizes feature engineering: momentum, moving averages, volatility measures, and volume-based indicators. These features are not magic; they are hypotheses you test in a disciplined way.

A strong takeaway is that signal design matters more than model complexity.

Supervised Learning in Finance

Classification and regression are used for directional bets, regime detection, and risk estimation. The biggest risks are overfitting, look-ahead bias, and survivorship bias.

The course highlights that a model that looks good in-sample is almost always misleading without careful validation.

Reinforcement Learning and Sequential Decisions

Trading is a sequential decision process. RL methods such as Q-learning provide a framework, but the course also shows why pure RL is difficult in finance: rewards are sparse, dynamics drift, and costs are real.

Risk Management and Evaluation

Performance is evaluated through risk-adjusted metrics:

• Sharpe and Sortino ratios
• Maximum drawdown
• Tracking error and information ratio

Backtesting is treated as a scientific experiment. If your evaluation is flawed, your strategy is too.

Practical Constraints

• Transaction costs can destroy a strategy that looks good on paper.
• Markets are adaptive; a good signal will decay.
• Data quality and cleaning are not optional.

Course Takeaways

• Models are only as good as their data and evaluation.
• Risk management must be built into the strategy, not added later.
• The edge is often in pipeline design, not in exotic models.

ML4T is a strong reminder that building a robust trading system requires discipline, skepticism, and a rigorous testing mindset.