Knightian Uncertainty¶

Introduction¶

Knightian uncertainty, named after economist Frank Knight (1921), represents a fundamental distinction between risk and uncertainty that has profound implications for quantitative finance and decision theory. While risk refers to situations where probability distributions are known or estimable, Knightian uncertainty captures scenarios where probabilities themselves are unknown or unknowable.

Knight's Risk-Uncertainty Distinction¶

1. Definitions¶

Risk: Situations where the probability distribution over outcomes is known or can be objectively estimated from data. This corresponds to the classical probabilistic framework where:

for a known probability measure \(P\).

Knightian Uncertainty (Ambiguity): Situations where the probability distribution is not uniquely specified, leading to a family of possible probability measures. Decision makers face model uncertainty about which probability measure governs the true data-generating process.

2. Historical Context and Motivation¶

Knight (1921) argued that true uncertainty—as opposed to measurable risk—is the source of entrepreneurial profit. In competitive markets, pure risk can be priced through actuarial methods, but genuine uncertainty cannot be fully hedged or insured against.

Key Insight: Economic agents often face situations where they have incomplete information about:

• The correct probability model
• Parameter values within a given model class
• The model structure itself (specification uncertainty)

Mathematical Framework¶

1. Savage's Subjective Expected Utility¶

Savage (1954) axiomatized decision-making under uncertainty using a unique subjective probability measure. However, this framework assumes that decision makers can always form precise probabilistic beliefs, which Knightian uncertainty challenges.

Savage's Axioms: Lead to a representation:

where \(u\) is a utility function and \(P\) is a unique subjective probability measure.

Limitation: Savage's framework cannot accommodate situations where the decision maker is genuinely uncertain about which probability measure to use.

2. Multiple Priors Representation¶

To capture Knightian uncertainty, we generalize from a single probability measure to a set of probability measures.

Definition (Set of Priors): Let \(\mathcal{P}\) be a convex, closed set of probability measures on \((\Omega, \mathcal{F})\). The decision maker evaluates acts \(f: \Omega \to \mathbb{R}\) using:

This is known as the maxmin expected utility representation (Gilboa-Schmeidler, 1989).

Interpretation: The decision maker:

1. Computes expected utility under each plausible probability measure in \(\mathcal{P}\)
2. Evaluates the act by its worst-case expected utility
3. Exhibits ambiguity aversion by focusing on the minimum

3. Variational Preferences¶

An alternative representation uses a penalty function for deviating from a reference measure \(P_0\).

Variational Representation (Maccheroni, Marinacci, Rustichini, 2006):

where \(c(P \| P_0)\) is a convex penalty function, often chosen as relative entropy:

with \(\theta > 0\) measuring the degree of ambiguity aversion.

Properties:

• When \(\theta \to 0\): Approaches classical expected utility with measure \(P_0\)
• When \(\theta \to \infty\): Approaches maxmin over \(\mathcal{P}\)
• Intermediate \(\theta\): Balances expected utility and model misspecification concerns

Ellsberg Paradox and Ambiguity Aversion¶

1. The Ellsberg Experiment¶

Ellsberg (1961) demonstrated systematic violations of Savage's axioms through the following thought experiment.

Setup: Two urns containing red and black balls:

• Urn 1: Contains 50 red and 50 black balls (known composition)
• Urn 2: Contains 100 balls of unknown composition (ambiguous)

Bet Structure:

• Bet A: Win $100 if red is drawn from Urn 1
• Bet B: Win $100 if black is drawn from Urn 1
• Bet C: Win $100 if red is drawn from Urn 2
• Bet D: Win $100 if black is drawn from Urn 2

Typical Preferences: Most people prefer:

• Bet A over Bet C (prefer known risk over ambiguity)
• Bet B over Bet D (prefer known risk over ambiguity)

Paradox: Under Savage's framework with subjective probability \(p\) for red in Urn 2:

• Preferring A over C implies: \(0.5 > p\)
• Preferring B over D implies: \(0.5 > 1-p\), which means \(p > 0.5\)

This contradiction demonstrates that people systematically avoid ambiguous situations, violating the independence axiom in Savage's theory.

2. Mathematical Analysis¶

Let \(P_{\text{Urn 2}}\) denote the decision maker's beliefs about Urn 2. The observed preferences suggest:

where \(\mathcal{P}_2\) is the set of priors for Urn 2.

This is consistent with maxmin preferences with:

Applications to Quantitative Finance¶

1. Model Uncertainty in Derivative Pricing¶

Recall (see § Robust No-Arbitrage Pricing and § Uncertain Volatility Models): under Knightian uncertainty about the data-generating measure (volatility, jump intensities, completeness), the single risk-neutral \(\mathbb{Q}\) is replaced by a set \(\mathcal{Q}\) of consistent martingale measures, and the robust price \(V_t = \inf_{\mathbb{Q} \in \mathcal{Q}} \mathbb{E}_{\mathbb{Q}}[e^{-r(T-t)}\Phi(S_T) \mid \mathcal{F}_t]\) generates a bid-ask spread \(V_{\text{sell}} < V_{\text{buy}}\) even in the absence of risk aversion.

2. Portfolio Selection under Ambiguity¶

Recall (see § Robust Portfolio Selection): the maxmin Markowitz problem \(\max_w \min_{(\mu, \Sigma) \in \Theta} \{w^\top \mu - (\lambda/2) w^\top \Sigma w\}\) with ellipsoidal mean uncertainty \(\Theta = \{(\mu,\Sigma): \|\mu - \hat{\mu}\|_{\Sigma^{-1}} \le \delta\}\) yields \(w^* = \Sigma^{-1}\hat{\mu}/[\lambda(1+\delta)]\), so ambiguity acts as an effective increase in risk aversion.

Theoretical Properties¶

1. Ambiguity Aversion vs Risk Aversion¶

Risk Aversion: Preference for certain outcomes over risky lotteries with the same expected value:

characterized by concavity of \(u\).

Ambiguity Aversion: Preference for known probability distributions over ambiguous ones:

where \(P_{\text{avg}} = \int_{\mathcal{P}} P \, d\nu(P)\) for some averaging measure \(\nu\).

Key Distinction:

• Risk aversion is about the curvature of utility over outcomes
• Ambiguity aversion is about preferences over probability measures

These are orthogonal concepts: one can be risk-neutral but ambiguity-averse.

2. Rectangularity and Dynamic Consistency¶

Recall (see § Dynamic Consistency): a set of priors \(\mathcal{P}\) is rectangular (Epstein-Schneider, 2003) if conditional distributions can be freely pasted across \(\mathcal{F}_t\) without leaving \(\mathcal{P}\); this is the necessary and sufficient condition for maxmin preferences to be dynamically consistent.

3. Connection to Robust Control¶

Recall (see § Multiplier Preferences): Hansen-Sargent (2001) robust control \(\min_u \max_w \mathbb{E}[\sum_t (x_t^\top Q x_t + u_t^\top R u_t - \theta w_t^\top w_t)]\) with linear dynamics \(x_{t+1} = Ax_t + Bu_t + Cw_t\) realizes Knightian uncertainty as a quadratic penalty on model distortions; the risk-sensitive limit \(\theta = -\beta\) recovers exponential utility \(V(x_0) = -\beta^{-1}\log \mathbb{E}[\exp(-\beta \sum_t r(x_t, u_t))]\).

Axiomatization and Representation Theorems¶

1. Gilboa-Schmeidler Axioms¶

Recall (see § Max-Min Expected Utility): Gilboa-Schmeidler (1989) characterize preferences satisfying weak order, continuity, monotonicity, uncertainty aversion, and certainty-independence/state-independence by

for a unique closed convex set of priors \(\mathcal{P}\) and utility \(u\) unique up to positive affine transformation.

2. Maccheroni-Marinacci-Rustichini Variational Preferences¶

Recall (see § Multiplier Preferences and § Entropy Penalization): MMR (2006) weakens GS independence to obtain the variational representation

with \(c\) convex and grounded. Special cases: \(c \equiv \infty\) off \(\{P_0\}\) gives Savage SEU; \(c \equiv 0\) on \(\mathcal{P}\), \(\infty\) otherwise gives Gilboa-Schmeidler maxmin; \(c(P) = \theta D_{\text{KL}}(P \| P_0)\) gives Hansen-Sargent multiplier preferences.

Calibration and Empirical Evidence¶

1. Measuring Ambiguity Aversion¶

Several experimental and market-based approaches measure the degree of ambiguity aversion.

Experimental Design: Compare choices between:

• Known probability bets (risk)
• Unknown probability bets (ambiguity)

Ambiguity Premium: The difference in certainty equivalents:

Empirical Findings:

• Most subjects exhibit \(\pi > 0\) (ambiguity aversion)
• Magnitude varies across individuals and contexts
• Professional traders often show less ambiguity aversion

2. Market Implications¶

Equity Premium Puzzle: Hansen-Sargent (2001) show that model uncertainty can help explain the historically high equity premium:

Options Market: The implied volatility smile may partially reflect ambiguity aversion about the true volatility process.

Bid-Ask Spreads: Wider spreads for less liquid assets may incorporate compensation for model uncertainty.

Connections to Information Theory¶

1. Relative Entropy and Model Distance¶

The relative entropy (Kullback-Leibler divergence) provides a natural measure of model distance:

Properties:

1. Non-negativity: \(D_{\text{KL}}(P \| Q) \geq 0\) with equality iff \(P = Q\)
2. Not symmetric: \(D_{\text{KL}}(P \| Q) \neq D_{\text{KL}}(Q \| P)\) in general
3. Convexity: \(D_{\text{KL}}(\cdot \| Q)\) is convex

Interpretation in Finance: \(D_{\text{KL}}(P \| Q)\) measures the "surprise" or information gained when the true model is \(P\) but we believed \(Q\).

2. Entropic Risk Measures¶

Recall (see § Coherent Risk Measures and § Entropy Penalization): the entropic risk measure \(\rho_\beta(X) = \beta^{-1}\log\mathbb{E}[e^{\beta X}]\) admits the variational dual \(\rho_\beta(X) = \sup_{\mathbb{Q} \ll \mathbb{P}}\{\mathbb{E}_\mathbb{Q}[X] - \beta^{-1} D_{\text{KL}}(\mathbb{Q}\|\mathbb{P})\}\), interpolating between expectation (\(\beta \to 0\)) and essential supremum (\(\beta \to \infty\)).

Advanced Topics¶

1. Second-Order Stochastic Dominance and Ambiguity¶

Definition (Second-Order Stochastic Dominance): Distribution \(F\) second-order stochastically dominates \(G\) (written \(F \succeq_{\text{SSD}} G\)) if:

Theorem: All risk-averse expected utility maximizers prefer \(F\) to \(G\) if and only if \(F \succeq_{\text{SSD}} G\).

Extension to Ambiguity: For multiple priors preferences:

for all concave \(u\) and convex \(\mathcal{P}\).

2. Smooth Ambiguity Model¶

Recall (see § Smooth Ambiguity Preferences): Klibanoff-Marinacci-Mukerji (2005) define \(V(f) = \int_\mathcal{P} \phi(\mathbb{E}_P[u(f)])\, d\mu(P)\), where \(u\) encodes risk aversion, \(\phi\) encodes ambiguity aversion, and \(\mu\) is a second-order prior, separating ambiguity from likelihood and yielding smoother preferences than maxmin.

3. Choquet Expected Utility¶

Recall (see § Choquet Expected Utility): for a convex capacity \(\nu\), the Choquet integral \(\int f\,d\nu = \int_0^\infty \nu(\{f \ge t\})\,dt\) coincides with \(\min_{P \in \text{core}(\nu)} \mathbb{E}_P[f]\) where \(\text{core}(\nu) = \{P: P(A) \ge \nu(A) \forall A\}\), linking non-additive integration to maxmin preferences.

Summary and Key Takeaways¶

1. Conceptual Distinction: Knightian uncertainty separates measurable risk (known probabilities) from genuine uncertainty (unknown probabilities)

2. Mathematical Frameworks: Multiple representation theorems capture ambiguity aversion:

3. Gilboa-Schmeidler maxmin expected utility
4. MMR variational preferences
5. Smooth ambiguity preferences

6. Choquet expected utility

7. Financial Applications: Model uncertainty affects:

8. Derivative pricing (bid-ask spreads)
9. Portfolio selection (conservative positions)
10. Risk management (stress testing)

11. Asset pricing puzzles (equity premium)

12. Dynamic Consistency: Rectangularity is necessary and sufficient for time-consistent preferences with multiple priors

13. Robust Control Connection: Knightian uncertainty provides decision-theoretic foundations for robust control in continuous-time finance

14. Measurement: Ambiguity aversion can be quantified through experimental elicitation and calibrated to market data

The study of Knightian uncertainty enriches our understanding of decision-making under incomplete information and provides robust approaches to financial modeling when model specification is uncertain.

Exercises¶

Exercise 1. In the Ellsberg paradox with Urn 2 containing 100 balls of unknown red/black composition, suppose a decision maker uses maxmin expected utility with the set of priors \(\mathcal{P}_2 = \{P : P(\text{Red}) \in [0.3, 0.7]\}\). Compute the maxmin expected utility of each bet (A, B, C, D) assuming linear utility \(u(x) = x\) and verify that the typical preference pattern (A over C, B over D) is reproduced.

Exercise 2. For the variational preference representation \(V(f) = \min_{P \in \mathcal{P}} \{\mathbb{E}_P[u(f)] + \theta D_{\text{KL}}(P \| P_0)\}\), show that the minimizing measure \(P^*\) satisfies

Interpret this result: what does the worst-case measure look like when the payoff \(f\) is high versus low?

Exercise 3. Consider a portfolio choice problem under Knightian uncertainty about expected returns. The mean return vector satisfies \(\mu \in \{\mu : \|\mu - \hat{\mu}\|_{\Sigma^{-1}} \leq \delta\}\). Derive the robust optimal portfolio

and show that increasing ambiguity \(\delta\) uniformly shrinks positions toward zero, equivalent to increasing the effective risk aversion.

Exercise 4. Prove that ambiguity aversion and risk aversion are orthogonal concepts by constructing an example of an agent who is risk-neutral (\(u(x) = x\)) but ambiguity-averse (uses maxmin over \(\mathcal{P}\)). Show that this agent's pricing kernel differs from the standard risk-neutral pricing kernel, and interpret the difference as an ambiguity premium.

Exercise 5. Explain the rectangularity condition for dynamic consistency of multiple priors. Consider a two-period model where \(\mathcal{P}\) contains two measures \(P_1\) and \(P_2\) on \(\{HH, HT, TH, TT\}\). Is the set \(\mathcal{P} = \{P_1, P_2\}\) rectangular if \(P_1(H_1) = 0.6\), \(P_1(H_2|H_1) = 0.5\), \(P_1(H_2|T_1) = 0.4\), and \(P_2(H_1) = 0.4\), \(P_2(H_2|H_1) = 0.5\), \(P_2(H_2|T_1) = 0.4\)? What if \(P_2(H_2|H_1) = 0.6\)?

Exercise 6. The smooth ambiguity model of Klibanoff-Marinacci-Mukerji uses the representation \(V(f) = \int_\mathcal{P} \phi(\mathbb{E}_P[u(f)]) \, d\mu(P)\). Consider the case where \(\phi(x) = -e^{-\alpha x}\) (exponential ambiguity aversion). Show that for normally distributed outcomes and a Gaussian second-order prior \(\mu\), the smooth ambiguity value reduces to a mean-variance criterion with an augmented variance term that reflects ambiguity.

Exercise 7. In the robust control formulation with linear dynamics \(x_{t+1} = Ax_t + Bu_t + Cw_t\) and quadratic costs, the worst-case disturbance is \(w_t^* = (1/\theta) C^\top V_{t+1} x_t\), where \(V_{t+1}\) is the value function matrix. For a scalar system with \(A = 0.9\), \(B = 1\), \(C = 0.5\), \(Q = 1\), \(R = 0.1\), solve for the robust optimal control and worst-case disturbance as a function of \(\theta\). What happens as \(\theta \to \infty\) (maximal robustness concern)?