Neural Architecture Search and Automated Machine Learning for Stock Prediction: Discovering Optimal Graph-Based Architectures for Financial Markets

Abstract

The design of neural network architectures for stock prediction has traditionally relied on manual experimentation and domain expertise, a process that is both time-consuming and potentially suboptimal. Different market conditions, asset classes, and prediction horizons may require different architectural configurations, yet the space of possible architectures remains largely unexplored. This paper proposes a Neural Architecture Search (NAS) framework tailored specifically for stock prediction, enabling the automated discovery of optimal graph-based neural network architectures without manual architecture engineering. Our approach builds upon the Stock State Space Graph architecture introduced by Lu, Hu, and Zhang, defining a novel search space that spans graph construction, message passing mechanisms, temporal aggregation, and state space configurations. We adapt explanation-based bias decoupling regularization principles from Zang and Liu's work on natural language inference to guide the architecture search toward configurations that are robust to overfitting and distribution shift. Through extensive experiments on benchmark financial datasets, we demonstrate that the NAS-discovered architectures substantially outperform manually designed baselines, achieving improvements of up to 18% in Sharpe ratio on the S&P 500 dataset. Furthermore, we show that the discovered architectures transfer effectively across different markets and asset classes, and that the architectures discovered for one prediction horizon can be adapted to others with minimal fine-tuning. Our work contributes to the growing body of research on AutoML in finance, providing a principled framework for automated architecture discovery that democratizes access to state-of-the-art prediction models.