How Artificial Intelligence Enriches the Traditional Data Science Toolbox in Finance

Department of Statistics, College of Literature, Science, and the Arts, University of Michigan-Ann Arbor, 48109 Ann Arbor, MI, USA

^* Corresponding author: chengcui@umich.edu

Abstract

With the continuous growth of the scale, heterogeneity and dynamics of financial data, traditional statistical and machine learning methods have gradually shown deficiencies in feature-engineering dependence, nonlinear modeling and stability in rapidly changing environments. Artificial intelligence (AI) has provided new tools and paradigms for financial data science, significantly enhancing performance in core tasks such as credit risk modeling, fraud detection, portfolio optimization, and market forecasting. Specifically, the three main lines of representation learning, relational learning, and time-series and real-time learning have enriched the modeling methods of data science, while decision engineering and governance & compliance have become the two major guarantees. However, in the process of implementation, AI still faces challenges such as data availability and out-of-distribution generalization/transferability, compliance-grade explainability and computational cost, probability calibration and decision engineering, dual sensitivity to data/structural drift and objective/hyperparameter design, large-scale deployment and latency, and the sustainability of returns and execution feasibility. This paper systematically reviews the expansion of the traditional data science toolkit by AI in the financial field and looks forward to the potential development paths in the future in areas such as financial foundation models, the integration of causal inference and constrained decision-making, and temporal/dynamic graphs and multi-agent market simulation.