Time-series forecasting under a low signal-to-noise ratio, using cross-sectional equity returns as the testbed. The goal is twofold: to investigate how well time-series foundation models transfer to high-noise settings, and to investigate deep learning architectures suited to such environments.

• Extending SLiDE, a Koopman-based recurrent model that adds exogenous inputs to SKOLR, for cross-sectional equity return prediction.
• Training it against a Sharpe ratio objective on PCA residuals, rather than a standard error metric.
• Evaluating time-series foundation models (Chronos, TabPFN) zero-shot under permutation testing.

Continuing through the 2026–2027 academic year.

Physical human–robot interaction in elderly care, where a robot and an older adult jointly carry an object and tremor makes naive force-following unsafe. Most HRI research relies on robot simulators or biomechanics platforms in isolation, with no coupled human–robot dynamics; this project builds a simulation framework that couples the two directly, so interaction forces emerge from the physics itself, and a controller on top of it for tremor mitigation.

• First-author extended abstract accepted at ICNR 2026; code released at NeuRoC-Lab/HRI-controller.
• Built a MuJoCo framework coupling a MyoSuite musculoskeletal arm to a torque-controlled Rethink Sawyer, so interaction forces emerge from the physics rather than being prescribed.
• Designed a series admittance–impedance controller reaching ~98% tremor attenuation at 4–14% tracking RMSE.

A short write-up is also on LinkedIn.