[Feature Request] Support Structured State Space Models (SSM, e.g. Mamba) in Qlib

[SYSUljz]

Motivation

Recently, Structured State Space Models (SSMs), such as S4, S5, and Mamba, have achieved strong performance in
long sequence modeling tasks. Compared to Transformers, SSMs can efficiently capture long-term dependencies
with linear time complexity, making them well-suited for large-scale time series forecasting in finance.

Proposal

I would like to suggest adding support for SSM-based deep learning models (starting with Mamba) in Qlib’s
model zoo. This will provide users with a new sequence modeling architecture in addition to existing RNN-
and Transformer-based models.

Benefits

• Improved efficiency in modeling long financial time series compared to Transformers.
• Strong scalability with lower computational overhead.
• Enriched model selection for researchers and practitioners using Qlib.

Possible Next Steps

• Implement a basic SSM/Mamba model class under qlib/model/.
• Provide training and inference examples for Alpha158/Alpha360 datasets.
• Add benchmark results to compare against Transformer-based models.

[whiteskylondon]

Very good idea