Generative Models for Trading

This unit explores generative and latent-variable models, representation and complexity reduction, training and scaling strategies, and multimodal data processing in trading and asset management contexts. Students will learn to preprocess and fuse market prices with audio/video, train and optimize generative models and LLMs (including FinBERT), and design hybrid solutions—such as RL‑based portfolio optimizers combined with generative/LLM components—for practical trading tasks.

Learning Objectives

• Analyze differences between discriminative and generative model families (e.g., VAEs, GANs, flows, LLMs) and justify selection of architectures for specific trading and asset-management problems based on representational capacity and complexity trade-offs
• Apply dimensionality reduction and latent-variable techniques to preprocess and fuse multimodal financial datasets (market prices, video, speech), producing correctly formatted inputs and baseline features for model training
• Demonstrate training, tuning, and scaling strategies for generative models and LLMs—addressing computational cost, inference latency, and emergent behavior—and optimize a FinBERT-like model for faster financial inference
• Evaluate and design hybrid model-driven trading solutions (e.g., deep RL portfolio optimizers integrated with generative models or LLM advisors), implement a proof-of-concept strategy, and assess its performance using financial metrics and risk measures