Model Risk in Preference-Learning Robo-Advisors

Abstract

Robo-advisors now manage roughly $2 trillion in assets globally, and the quality of their investment advice rests on correctly learning each client’s risk preferences. The authors provide the first comprehensive risk management framework for algorithmic preference learning. Using a nonparametric utility reconstruction algorithm, they show that a robo-advisor can faithfully infer an investor’s risk preferences from observed trading behavior and derive near-optimal portfolio allocations across a wide range of wealth levels. Systematic stress testing reveals that the robo-advisor is robust to moderate behavioral noise. To account for investor’s preference shifts and behavioral noise, the authors introduce a two-layer automated circuit breaker that detects most failure modes in real time and triggers human oversight before severe wealth losses occur. They propose targeted mitigation strategies that reduce realized portfolio allocation errors for investors, regardless of their degree of supervision by the robo-advisors.