Most introductory courses stop at linear models. This program picks up where they leave off. Support Vector Machines let you find nonlinear decision boundaries in high-dimensional feature spaces. That matters when you're trying to predict market movements based on dozens of technical indicators and fundamental ratios.

Why SVMs for Trading

SVMs excel at binary classification, which maps perfectly to trading decisions: buy or don't buy, enter or exit. The kernel trick lets you capture complex interactions between features without manually engineering every possible combination. We cover RBF, polynomial, and linear kernels, explaining when each makes sense for financial data.

You'll implement a pairs trading strategy using SVMs to identify when two correlated stocks diverge enough to signal a trade. The model learns from historical spread behavior and generates entry and exit signals. We discuss transaction costs, slippage, and why most backtests overestimate real performance.

Neural Networks for Pattern Recognition

The second half introduces feedforward neural networks for time-series prediction. These aren't deep learning models with millions of parameters. We focus on shallow networks with one or two hidden layers that can learn nonlinear relationships in market data without overfitting.

You'll build a network that predicts next-day returns using a combination of price patterns, volume data, and sentiment indicators. The program covers activation functions, backpropagation mechanics, and regularization techniques. You'll understand why adding more layers doesn't always improve financial predictions.

Real Performance Metrics

Every model gets evaluated on out-of-sample data using walk-forward analysis. We calculate information ratio, maximum drawdown, and win rate. You'll see how models that look great in training often fail in live testing. The program emphasizes realistic expectations about what machine learning can and cannot do in financial markets.

We use Python with scikit-learn for SVMs and TensorFlow for neural networks. You'll write vectorized code that runs efficiently on historical data and learn to structure experiments so you can iterate quickly on model architectures.