Learning to Manage Investment Portfolios beyond Simple Utility Functions

Our paper, Learning to Manage Investment Portfolios beyond Simple Utility Functions, was presented at the ACM International Conference on AI in Finance (ICAIF) 2025.

Read it on arXiv · ACM Digital Library

It is joint work with Mahmoud Mahfouz, Anisoara Calinescu, and J. Doyne Farmer. Fund managers optimise for competing goals that a single utility function cannot capture. Instead of specifying one, we learn a latent representation of each manager’s strategy directly from the data — modelling the portfolio weights conditional on stock characteristics, past returns, and previous holdings, with a GAN trained on observed holdings rather than labelled objectives.

An informal, longer treatment is in the thesis chapter Imitating Fund Managers using GANs. The model and results are in the paper.