Short-Term Stock Portfolio Optimization Based on LSTM-GNN Hybrid Models

Sammanfattning

Introduction: The integration of artificial intelligence in quantitative finance traditionally relies on purely data-driven deep learning models. However, in low signal-to-noise environments like the China A-share market, these conventional architectures frequently struggle with spurious correlations and structural overfitting. This thesis explores a shift from implicit statistical learning to a knowledge-infused approach by explicitly embedding macroeconomic fundamental relationships into neural networks. Research Question: The primary research question is: How can an end-to-end (E2E) knowledge-infused spatio-temporal framework (LSTM-GNN) be constructed to mitigate low SNR challenges and capture non-Euclidean asset correlations, while maintaining robust portfolio optimization and economic interpretability in the China A-share market? Methodology: To answer this question, a hybrid spatio-temporal architecture (LSTM-GAT) was developed and evaluated using a decade of China A-share market data (2016-2025). A Long Short-Term Memory (LSTM) module processed 20-day historical time-series sequences to capture temporal momentum. Concurrently, a multi-head Graph Attention Network (GAT) routed cross-sectional signals guided by a supply chain graph, which served as a macroeconomic proxy. The model was benchmarked against a data-driven Pearson correlation baseline and optimized directly using a differentiable negative Sharpe ratio loss. Results: The knowledge-infused hybrid model achieved an out-of-sample annualized return of 24.03\% and a Sharpe Ratio of 1.25, significantly outperforming the Pearson correlation baseline (Sharpe Ratio of 0.12). Ablation studies confirmed that both temporal memory and spatial routing are necessary to maximize risk-adjusted returns. Furthermore, extracting the GAT attention coefficients successfully visualized capital spillover effects across diverse industrial sectors. Discussion: These findings demonstrate that embedding explicit economic rationale acts as a strong inductive bias, enabling the model to filter out transient market noise. While the study is primarily limited by the use of a static graph topology and potential concept drift, it offers a deployable framework with interpretable allocation logic, aligning with the principles of transparent quantitative portfolio optimization.