Learning from noisy market environments.

The question

Can predictive and reinforcement-learning systems extract useful signal without relying on unrealistic execution assumptions?

The approach

Build multi-asset data infrastructure and evaluate recurrent, convolutional, attention-based, Transformer, and tree-based architectures. Separate representation, prediction, decision, and execution; compare against deterministic quantitative strategies.

Validation focus

• Include commissions, slippage, rollover, sessions, and DST handling.
• Enforce non-lookahead logic, non-repainting signals, and consistent OCO execution.
• Evaluate regime dependence, out-of-sample stability, and cross-market behavior.