End-to-end learning of optimal portfolios

keywords: machine learning, decision making, portfolio optimization, generalization, risk-awareness, transparency, non-ergodic processes

abstract: We propose to develop end-to-end machine learning methods suited for structured problems, where the learning utility is given by performance within an external optimization task, rather than accuracy of the model itself. Particularly, we target a well-studied optimization task of allocating a distribution of resources over a, possibly structured, set of stochastic assets, subject to diverse criteria of expected utility, risk and transparency. We intend to integrate developments in symbolic, mathematical, and differentiable programming, to allow for a transparent end-to-end learning process across variety of the portfolio optimization formulations. Following our successful preliminary experiments against state-of-the-art in sports prediction markets, we plan to extend towards full end-to-end setting to prove advantage of the proposed ideas over the standard methods approaching prediction and subsequent portfolio optimization separately in real benchmarks. We will address corresponding issues of generalization, tractability, and use of utility theory in the proposed learning setting.