The rapid advancement of artificial intelligence (AI) has reshaped numerous fields, and finance is no exception. Asset pricing, a domain traditionally dominated by linear models and factor-based approaches, is now experiencing a transformative shift due to AI’s capacity to uncover persistent predictive patterns in financial data. The introduction of large-scale AI models – particularly transformer-based architectures – has significantly enhanced our capability to model complex relationships among assets and firms (Eisfeldt et al. 2023), leading to improved forecasting and risk assessment.

In this column, we explore how AI-driven asset pricing models, particularly those incorporating transformers, leverage cross-asset information sharing to reduce pricing errors and enhance predictive accuracy. These innovations offer a novel perspective on how financial markets process information and determine asset prices.

Traditional versus AI-driven asset pricing

Most traditional asset pricing models, such as the Fama and French (1993) framework, rely on predefined factors that explain asset returns. While effective, these models assume a fixed and linear relationship between asset characteristics and expected returns. AI, on the other hand, introduces a non-linear, data-driven approach, identifying patterns that are often invisible to traditional methods.

Machine learning models, including tree-based methods and neural networks (Gu et al. 2020), have improved asset return prediction by capturing complex relationships between firm characteristics and returns. However, these models typically focus on ‘own-asset prediction’, meaning that they use only an asset’s individual characteristics to estimate future returns. This approach ignores the broader context in which assets interact.

The role of transformers in asset pricing

In a recent paper (Kelly et al. 2024), we introduce the Artificial Intelligence Pricing Model (AIPM), which embeds transformer networks into the stochastic discount factor (SDF) framework. While initially developed for natural language processing, we show that transformers are remarkably effective in financial applications due to their ability to capture cross-sectional dependencies across assets.

Unlike traditional machine learning models, transformers incorporate the ‘attention mechanism’, allowing them to dynamically adjust the weight placed on different inputs based on their relevance. In the context of asset pricing, this means that the model not only considers an asset’s own characteristics but also how these characteristics interact with those of other assets. This approach significantly enhances predictive power by leveraging market-wide information.

Empirical findings: Performance of AI-based models

We evaluated the performance of the transformer-based AIPM using a dataset of US stock returns and conditioning variables. Compared to traditional asset pricing models and other machine learning approaches, the AIPM demonstrated:

• Lower pricing errors: The model achieved significantly smaller out-of-sample pricing errors compared to traditional factor-based models and neural networks without attention mechanisms.
• Higher Sharpe ratios: By integrating cross-asset dependencies, the transformer-based model outperformed existing approaches in terms of risk-adjusted returns out-of-sample.
• Scalability and complexity gains: We found that increasing model complexity – by incorporating deeper transformer layers – consistently improved predictive performance. This supports the notion that AI models benefit from higher parameterisation when applied to asset pricing.

Empirical evidence from the literature on large language models (e.g. Kaplan et al. 2020) indicates that adding more transformer blocks enhances the capacity of models to effectively represent language. Models with deeper architectures are able to capture more abstract features and longer-range dependencies than shallower models. Each additional layer refines the attention distributions, allowing the model to consider both short-term and long-term relationships. Interestingly, the same benefits of transformer complexity emerge in the context of the asset pricing model.

These findings suggest that AI-driven asset pricing models are more efficient in processing vast amounts of financial data, leading to more accurate and robust predictions.

Implications for investors and policymakers

The integration of AI into asset pricing has profound implications for market participants and policymakers. Investors can benefit from improved portfolio allocation strategies driven by AI’s capability to identify subtle pricing inefficiencies, a concept aligned with research on AI and personality traits shaping economic returns (Makridis 2025). Meanwhile, regulators and policymakers must consider the impact of AI-driven trading on market stability and efficiency.

Moreover, the adoption of AI in asset pricing challenges traditional views on market efficiency, echoing broader concerns about AI’s macroeconomic impact and productivity gains (Filippucci et al. 2024). If AI models consistently outperform classical frameworks, it may suggest that markets are less efficient than previously thought, opening the door for further research into the nature of pricing anomalies.

Conclusion

The fusion of artificial intelligence and finance is revolutionising asset pricing. Our research demonstrates that transformer-based models significantly enhance return predictions by leveraging cross-asset information sharing. As AI continues to evolve, its role in financial decision-making will only grow, offering new opportunities for investors and reshaping our understanding of market dynamics.

References

Gu, S, B Kelly and D Xiu (2020), “Empirical Asset Pricing via Machine Learning”, Review of Financial Studies 33 (5): 2223-2273.

Eisfeldt, A, G Schubert and M B Zhang (2023), “Generative AI and Firm Valuation”, VoxEU.org, 4 June.

Fama, E and K French (1993), “Common risk factors in the returns on stocks and bonds”, Journal of Financial Economics 33(1): 3-56.

Filippucci, F, P Gal, and M Schief (2024), “Miracle or Myth? Assessing the Macroeconomic Productivity Gains from Artificial Intelligence”, VoxEU.org, 8 December.

Kaplan, J, S McCandlish, T Henighan et al. (2020), “Scaling laws for neural language models”, arXiv preprint arXiv:2001.08361.

Kelly, B Kuznetsov, S Malamud and T Xu (2024), “Artificial Intelligence Asset Pricing Models”, NBER Working Paper No. w33351

Makridis, C (2025), “The Role of Personality Traits in Shaping Economic Returns Amid Technological Change”, VoxEU.org, 31 January.

Vaswani, A, N Shazeer, N Parmar, J Uszkoreit, L Jones, A N Gomez, L Kaiser and I Polosukhin (2017), “Attention is all you need”, Advances in Neural Information Processing Systems 30.