Jiajing Sun

Jiajing Sun

PhD, CFA charterholder, FIMA, Member of the London Mathematical Society. Associate Professor and the Deputy Director of the Department of Statistics and Data Science, School of Economics and Management, University of the Chinese Academy of Sciences.

Testing Generative Artificial Intelligence for Economic Time Series

Published:

Status: Working paper. Submitted to the Journal of Time Series Analysis special issue on artificial intelligence in macroeconomics and finance.

This paper studies whether synthetic time series generated by artificial intelligence can reliably stand in for observed economic and financial data. Rather than relying on visual comparisons, summary statistics, or model rankings, it formulates synthetic-data validation as a formal finite-horizon specification test.

Authors:

Yongmiao Hong, Oliver Linton, Jiajing Sun, and Abderrahim Taamouti.

Summary and Contribution:

The test compares overlapping blocks from a held-out observed series with blocks from multiple independently generated paths using characteristic-kernel maximum mean discrepancy. The null hypothesis asks whether the two samples share the same finite-horizon joint distribution, covering marginal behavior, serial dynamics, cross-series dependence, lead-lag relationships, and nonlinear or tail features.

The two samples have different dependence structures: observed blocks overlap and remain serially dependent, while synthetic observations are clustered within generated paths. The proposed bootstrap therefore combines dependent multipliers for the observed series with cluster multipliers for the generated paths. Conditional on a fitted generator, the paper establishes bootstrap validity and asymptotically correct size.

Evidence and Application:

In 2,000-replication simulations, the proposed procedure delivers rejection rates close to the nominal 5% level under the exact null. By contrast, omitting uncertainty from the observed series produces rejection rates close to 50%, demonstrating the importance of calibrating both sources of dependence.

The empirical application evaluates three fitted generators using six Federal Reserve H.10 exchange rates. All three generators are rejected on the 2018-2020 held-out sample, and diagnostic analysis shows that no single generator dominates across marginal, serial, cross-currency, and tail characteristics.

Keywords:

Generative artificial intelligence; empirical characteristic function; maximum mean discrepancy; specification testing; dependent multiplier bootstrap; synthetic time series.

Availability:

The manuscript is not hosted on this website because it is currently under journal consideration.

Recommended citation: Hong, Y., Linton, O., Sun, J., & Taamouti, A. (2026). Testing Generative Artificial Intelligence for Economic Time Series. Working paper.