Weather sentiment driven quantum graph networks for intraday stock prediction

Abstract

Intraday stock market movements are highly sensitive to external factors, including weather-induced investor sentiment, making accurate prediction a significant challenge for financial analysts and automated trading systems. This study proposes an Efficient Intraday Analysis on Stock Market Behavior and Sentiment Affected by Weather using Hamiltonian Graph Quantum Dandelion Sample Generative Aggregate Adversarial Attention Networks (Ham-GQDS-G3AN). The prediction task is formulated as a three-class directional classification problem —upward trend, downward trend, and stable market — based on hourly next-return labels constructed using a threshold-based scheme (δ = 0.02% × MAD). The model is evaluated using a strict chronological train–test split (80% training: 44,027 h; 20% testing: 11,007 h) with rolling-window validation (window=2000 h, step=500 h, ∼85 folds) to prevent look-ahead bias. This data includes all intraday weather, sentiment and stock market observations from the KOSPI market (June 2000 – December 2021) with 55,034 observations per hour ranging from 9:00-16:00. Data preprocessing uses only a subset of the training data to avoid data leakage and calculates the Correlation Coefficients and Min–Max Normalization (CCMN) only on this subset. Feature extraction is done with the Kolmogorov-Arnold Vision Transformer (KAVT), which is also calculated in the training set. The Ham-GQDS-G3AN framework is a combination of Hamiltonian Quantum Generative Adversarial Networks (HQuGANs), Graph Self-Attention Network (GSAN), and optimization using Dandelion Optimizer Algorithm (DOA). The underlying causal mechanism is based on the so-called ‘mood-misattribution hypothesis’ that suggests that weather influences investor psychological disposition, leading to systematic errors in the perception of risk and trading behavior. The proposed approach shows 99.98% classification accuracy which is validated by confusion matrix, ROC-AUC ( macro = 0.9996), and Diebold–Mariano statistical significance tests against seven competing baselines.