Definition
Ambiguity aversion (uncertainty aversion, Knightian uncertainty) is the decision-theoretic stance that a decision maker facing unknown probabilities — not just known risk — prefers acts that are robust across a set of possible probability models rather than committing to one. Formally it is captured by the multi-prior / maxmin expected utility criterion (Gilboa-Schmeidler): maximize the worst-case expected utility over a set of priors. Garlappi, Uppal & Wang (2007) apply it to portfolio choice to handle uncertainty about estimated expected returns (Garlappi-Uppal-Wang 2007).
Key Ideas
- Risk vs. ambiguity. Classical expected utility treats a single known probability distribution (risk). Ambiguity is uncertainty about that distribution itself — the Ellsberg-paradox phenomenon that people avoid bets with unknown odds.
- Multi-prior / maxmin. Rather than a single prior (Bayesian) the agent entertains a set of priors and, being averse to ambiguity, evaluates each act by its minimum expected utility over that set, then maximizes — maxxminμ∈PEμ[U(x)].
- Bayesian and neutral as a special case. The standard Bayesian approach to estimation error uses a single prior and is uncertainty-neutral (it averages over the prior); the multi-prior model nests it and adds explicit aversion via the inner minimization.
- Confidence-interval priors (Garlappi-Uppal-Wang). In their portfolio application the set of priors is a confidence interval/region around the estimated expected return μ^; aversion is a minimization over that region. The size of the region can be common to all assets, or differ across assets/subsets.
- Shrinkage interpretation. In several special cases the optimal multi-prior portfolio has a closed form: a shrinkage of the mean-variance portfolio toward the risk-free asset or the minimum-variance portfolio — the more ambiguity, the stronger the shrinkage toward the ambiguity-robust benchmark.
Why It Matters
- A third response to mean-variance instability. Where robust optimization worst-cases over an uncertainty set and Bayesian shrinkage averages over a single posterior, ambiguity aversion worst-cases over a set of priors — a decision-theoretically grounded middle ground that reduces to the Bayesian case when aversion is switched off.
- Stability and out-of-sample performance. Accounting for parameter uncertainty this way reduces the fluctuation of portfolio weights over time and improved out-of-sample performance in the authors' data — directly addressing the estimation-risk problem.
- Reach beyond portfolios. The maxmin/multi-prior apparatus (and its robust-control cousin, Hansen-Sargent) applies wherever model misspecification must be hedged decision-theoretically.
Open Questions
- Calibrating the set of priors (the confidence level / ambiguity size) is the crux — too large is uselessly conservative, too small is not robust.
- Relationship between maxmin multi-prior, smooth ambiguity (Klibanoff-Marinacci-Mukerji), and robust control — when they coincide and when they differ.
- Extending closed-form shrinkage results to covariance-matrix ambiguity and dynamic/multi-period settings.
Related