Asset pricing theory rests on the principle that differences in expected returns across assets are driven by their exposures to systematic risk factors. Identifying the ''right'' factors-whether observable or latent-remains a central challenge in empirical finance. Traditional latent factor models offer a parsimonious framework for summarizing information from hundreds of observable firm characteristics; however, they are typically estimated solely from return matrices, which limits their ability to capture time-varying, firm-specific dynamics. This study proposes a novel framework---Neural Instrumented Factorization (NeurIF)---that leverages firm characteristics as instruments to learn economically meaningful and time-varying latent factors. NeurIF integrates spatial and temporal attention mechanism to capture nonlinear relationships between firm characteristics and asset returns, jointly learning both the latent factors and their dynamic loadings. The model incorporates orthogonality constraints and deviation-based penalties to ensure the interpretability and alignment of latent factors with observed firm characteristics. Empirical evaluations on real-world asset pricing data reveal that NeurIF consistently outperforms several state-of-the-art transformer based models in return prediction, with improvements ranging from 1% to 18% in test data. Furthermore, the learned factor loadings can generate statistically significant long-short portfolio returns and are not subsumed by other observable factors. The embedded latent factors also exhibit strong explanatory power across several cross-sectional asset pricing anomalies, highlighting their economic relevance and robustness.

Recent studies suggest that financial networks play an essential role in asset valuation and investment decisions. Unlike road networks, financial networks are neither given nor static, posing significant challenges in learning meaningful networks and promoting their applications in price prediction. In this paper, we first apply the attention mechanism to connect the "dots" (firms) and learn dynamic network structures among stocks over time. Next, the end-to-end graph neural networks pipeline diffuses and propagates the firms' accounting fundamentals into the learned networks and ultimately predicts stock future returns. The proposed model reduces the prediction errors by 6% compared to the state-of-the-art models. Our results are robust with different assessment measures. We also show that portfolios based on our model outperform the S&P-500 index by 34% in terms of Sharpe Ratio, suggesting that our model is better at capturing the dynamic inter-connection among firms and identifying stocks with fast recovery from major events. Further investigation on the learned networks reveals that the network structure aligns closely with the market conditions. Finally, with an ablation study, we investigate different alternative versions of our model and the contribution of each component.