Koopman Dynamics for Portfolio Rebalancing

Learning sparse Koopman representations to improve long-horizon portfolio dynamics modeling and rebalancing.

This project studies Koopman-based latent dynamics modeling for financial time series and portfolio control. The codebase implements sparse Koopman autoencoder variants to better model nonlinear market dynamics while preserving stable long-horizon prediction behavior.

Summary

  • Implements multiple Koopman autoencoder variants (including sparse and LISTA-based encoders)
  • Evaluates long-horizon rollout quality under several re-encoding strategies
  • Supports reproducible training/evaluation pipelines and structured metrics exports

Poster

Koopman dynamics poster preview