AceMeistr/Liquid-NN-Boosting-Sequential-Model
Implementation of Liquid Neural Networks with Boosters using Sequential Architecture
What's novel
Implementation of Liquid Neural Networks with Boosters using Sequential Architecture
Code Analysis
8 files read · 3 roundsA multi-stage market index prediction system that combines a Mamba/SSM-based Liquid Neural Network with XGBoost residual modeling, spectral correlation eigenvalue features, temporal graph convolutions on constituent networks, and Bates-Granger/Hurst-exponent regime-adaptive ensemble blending, all tu
Strengths
Genuinely deep implementation with novel architectural choices: a pure-PyTorch Mamba fallback, spectral eigenvalue features from rolling correlation matrices, latent-space feature transfer from LNN to XGBoost, and Bates-Granger optimal weighting with Hurst-exponent regime detection — none of which are standard boilerplate. The data pipeline is well-engineered with vectorized target computation, three leakage-safe split strategies, and production-grade Optuna tuning with PBO penalties and thread-
Weaknesses
The README is vague and undersells the actual complexity — it mentions 'ODEs' but the code uses Mamba/SSM, and the '3-stage' claim is only partially reflected in the code structure. Test coverage is thin relative to the codebase size (132 source files vs ~10 test files), and several critical paths like the hybrid training loop, graph training, and Optuna objective were not fully verifiable due to file size.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
10
Contributors
1
Files
108
Active weeks
4
Repository
Language
Python
Stars
1
Forks
0
License
Apache-2.0