Start with how neural networks actually work. Not just the math—though we cover that—but the intuition behind why certain architectures perform better for financial data. You'll build your first model by week two.

We focus on understanding gradient descent, backpropagation, and optimization techniques specifically for time-series prediction. The assignments use historical price data so you can see immediately whether your model works.

Financial data has patterns that normal datasets don't have. Autocorrelation, seasonality, regime changes. LSTMs and transformers handle these differently, and you need to know when to use which.

We spend three weeks on this because getting it wrong is expensive. You'll work with tick data, daily closes, and everything in between. The final project involves predicting volatility shifts—notoriously difficult but good practice.

Theory meets reality here. You'll use reinforcement learning for asset allocation—watching agents learn to balance risk and return through thousands of simulated trading days.

This module gets intense. We cover Q-learning, policy gradients, and actor-critic methods. But more importantly, you'll understand why these approaches sometimes fail spectacularly and how to recognize those situations early.

News sentiment, satellite imagery, transaction data—markets respond to information from everywhere now. This module teaches you how to incorporate non-traditional signals into prediction models.

You'll work with natural language processing for earnings calls and social media. The challenge is filtering signal from noise, which is harder than it sounds when Twitter trends can move markets.

A model that works on your laptop isn't useful until it runs in production. This final module covers deployment, monitoring, and maintenance of financial ML systems.

You'll learn about latency constraints, model versioning, and graceful degradation. We also cover the painful topic of explaining black-box predictions to compliance teams—because you'll need to do that.