[Paper Review] Learning Risk Preferences from Investment Portfolios Using Inverse Optimization

Shi Ming Yu, Haoran Wang+1 more|arXiv (Cornell University)|Oct 4, 2020

Financial Markets and Investment Strategies4 citations

TL;DR

This paper proposes an inverse optimization framework to learn time-varying risk preferences from observed investment portfolios and market data, using mean-variance portfolio theory as the forward model. By inferring the risk preference parameter that best explains actual portfolio allocations, the method enables real-time, data-driven estimation of investor risk tolerance, validated against established risk metrics and showing strong consistency with market benchmarks.

ABSTRACT

The fundamental principle in Modern Portfolio Theory (MPT) is based on the quantification of the portfolio's risk related to performance. Although MPT has made huge impacts on the investment world and prompted the success and prevalence of passive investing, it still has shortcomings in real-world applications. One of the main challenges is that the level of risk an investor can endure, known as \emph{risk-preference}, is a subjective choice that is tightly related to psychology and behavioral science in decision making. This paper presents a novel approach of measuring risk preference from existing portfolios using inverse optimization on the mean-variance portfolio allocation framework. Our approach allows the learner to continuously estimate real-time risk preferences using concurrent observed portfolios and market price data. We demonstrate our methods on real market data that consists of 20 years of asset pricing and 10 years of mutual fund portfolio holdings. Moreover, the quantified risk preference parameters are validated with two well-known risk measurements currently applied in the field. The proposed methods could lead to practical and fruitful innovations in automated/personalized portfolio management, such as Robo-advising, to augment financial advisors' decision intelligence in a long-term investment horizon.

Motivation & Objective

• To address the limitation of static, survey-based risk preference estimation in automated investing by learning risk preferences directly from observed portfolio allocations.
• To develop a dynamic, real-time method for estimating time-varying risk preferences using concurrent market and portfolio data.
• To validate the inferred risk preferences against established financial risk measures such as beta and inverse Sharpe ratios.
• To enable practical deployment in automated portfolio management systems by providing actionable, continuously updated risk tolerance parameters.
• To demonstrate the method’s robustness and consistency using robotic investment portfolios and real mutual fund holdings over 10–20 years of data.

Proposed method

• Formulate the portfolio allocation decision as a mean-variance optimization problem where risk preference is a learnable parameter.
• Apply inverse optimization to infer the risk preference parameter that would have generated the observed portfolio allocations, given market returns and asset prices.
• Use a convex optimization framework to solve for the risk preference parameter that minimizes the duality gap between observed and optimal portfolios.
• Extend the method to sector and factor-level allocations to assess risk preference across different investment dimensions.
• Incorporate time-series data to estimate risk preferences dynamically, allowing for inter-temporal variation in investor risk tolerance.
• Validate results using benchmark risk measures such as CAPM beta and inverse Sharpe ratios to ensure financial intuition and consistency.

Experimental results

Research questions

• RQ1 Can risk preferences be accurately inferred from observed portfolio allocations using inverse optimization, without relying on surveys or hypothetical scenarios?
• RQ2 How do the inferred risk preference parameters compare to established financial risk metrics such as beta and inverse Sharpe ratios?
• RQ3 To what extent do risk preferences estimated from real-world portfolios exhibit time-varying behavior, and how stable are they across different investment styles or sectors?
• RQ4 Can the inferred risk preferences be used as reliable inputs for automated, personalized portfolio construction in Robo-advisory systems?
• RQ5 How does the method perform when applied to both simulated robotic portfolios and real mutual fund holdings over long time horizons?

Key findings

• The proposed inverse optimization method successfully infers time-varying risk preferences from observed portfolios, with results consistent across multiple validation metrics.
• Estimated risk tolerance values for S&P 500 benchmark funds ranged between 0.2 and 0.5 in sector space, reflecting moderate risk exposure.
• Actively managed funds like VSEQX and VSTCX exhibited higher risk tolerance values (above 0.5), aligning with their higher volatility and active management style.
• The order of estimated risk tolerance values across mutual funds showed strong consistency with their CAPM beta values, with most scatter plots near the diagonal in Figure 4(h).
• Inverse Sharpe ratios in factor space were smoother and more stable than in sector space, where volatility and wider confidence intervals were observed.
• The method enables real-time, data-driven risk preference estimation, making it suitable for dynamic, personalized Robo-advisory systems that update risk parameters as market conditions evolve.