Source-linked AI summary
A Unified Framework for Dynamic Pari-Mutuel Information Market Design
Shipra Agrawal, Erick Delage, Mark Peters, Zizhuo Wang, Yinyu Ye
TL;DR
Contingent-claim markets face liquidity and mechanism-design challenges, while existing pari-mutuel market makers arise from different formulations. The paper develops a generalized SCPM convex framework that unifies these mechanisms and connects market organization to risk minimization. It establishes truthful pricing, efficient computation, proper scoring, and worst-case-loss guarantees, while clarifying market-maker risk attitudes.
Problem
Contingent-claim markets can face liquidity problems, and existing pari-mutuel mechanisms have differing scoring-rule, cost-function, and convex-optimization foundations.
Method
The paper generalizes SCPM into a unified convex optimization framework and uses cost functions and duality to relate market mechanisms to risk minimization.
Results
The framework provides myopically truthful pricing, efficient convex cost-function computation, proper scoring, and worst-case-loss guarantees while representing existing market makers.
Takeaways & Limitations
Mechanisms can be compared and designed through their utility functions, which determine properties such as scoring properness, loss bounds, and market-maker risk attitude.
Takeaways & Limitations
The framework requires a technical feasibility and boundedness condition on the utility function for every nonnegative position vector.
Abstract
from arXiv · showhide
Recently, several new pari-mutuel mechanisms have been introduced to organize markets for contingent claims. Hanson introduced a market maker derived from the logarithmic scoring rule, and later Chen and Pennock developed a cost function formulation for the market maker. On the other hand, the SCPM model of Peters et al. is based on ideas from a call auction setting using a convex optimization model. In this work, we develop a unified framework that bridges these seemingly unrelated models for centrally organizing contingent claim markets. The framework, developed as a generalization of the SCPM, will support many desirable properties such as proper scoring, truthful bidding (in a myopic sense), efficient computation, and guarantees on worst case loss. In fact, our unified framework will allow us to express various proper scoring rules, existing or new, from classical utility functions in a convex optimization problem representing the market organizer. Additionally, we utilize concepts from duality to show that the market model is equivalent to a risk minimization problem where a convex risk measure is employed. This will allow us to more clearly understand the differences in the risk attitudes adopted by various mechanisms, and particularly deepen our intuition about popular mechanisms like Hanson's market-maker. In aggregate, we believe this work advances our understanding of the objectives that the market organizer is optimizing in popular pari-mutuel mechanisms by recasting them into one unified framework.
1. INTRODUCTION
Contingent-claim markets can aggregate information but may face liquidity problems, motivating automated market makers and renewed study of pari-mutuel mechanisms. The paper proposes a unified convex framework that connects existing mechanisms and supports desirable market-design properties.
- Motivation: Prediction markets aggregate information about events, while contingent-claim markets also support hedging and entertainment purposes.Reported applications include election markets and markets for event outcomes such as sports results or economic indicators.
- Motivation: Thin markets can suffer liquidity problems, leading organizers to consider automated market makers and their risk tolerance.The organizer must choose pricing rules while deciding how much risk to accept.
- Existing mechanisms: Existing mechanisms have different origins: SCPM uses convex optimization, whereas Hanson’s LMSR derives pricing from the logarithmic scoring rule.Chen and Pennock also provided a cost-function formulation for market scoring-rule mechanisms.
- Contributions: The paper proposes a generalized SCPM as a unified convex optimization framework for market makers.The framework is intended to connect SCPM, market scoring rules, and cost-function-based markets.
- Contributions: The framework provides truthful pricing, efficient cost-function formulation, scoring properness, and worst-case-loss guarantees.These properties are presented as desirable characteristics of the new model.
- Contributions: Any market maker based on a proper scoring rule can be represented within the SCPM model, while suitable framework objectives correspond to proper scoring rules.The converse holds under easily verifiable conditions, establishing the claimed unification.
- Contributions: The framework’s market-maker model is equivalent to convex risk minimization, linking mechanism design choices to the organizer’s risk attitude.The paper uses this equivalence to interpret mechanisms such as LMSR.
2. BACKGROUND
The background introduces market scoring rules, cost-function market makers, and SCPM as alternative ways to price contingent claims. These mechanisms use scoring, convex costs, or optimization-based order acceptance to determine prices and manage organizer risk.
- Market setup: A contingent claim pays $1 when its specified outcome occurs, and traders submit orders over mutually exclusive outcomes.The market maker determines the price charged for each new order.
- Market Scoring Rules: A proper scoring rule motivates truthful probability reporting, and Hanson’s MSR charges traders for changing the current probability estimate.The trader pays the score associated with the old estimate and receives the score associated with the new estimate.
- Market Scoring Rules: LMSR and quadratic scoring are examples of market scoring rules, while LMSR is known to elicit truthful bids.Hanson’s MSR is also described as a pari-mutuel automated market maker with bounded organizer risk.
- Cost-function market makers: In a cost-function market maker, an order a costs C(q + a) − C(q), and the infinitesimal price of outcome i is ∂C/∂q_i.The vector q records the claims currently held by traders.
- Cost-function market makers: Chen and Pennock showed that scoring rules can have equivalent cost-function formulations, enabling later equivalence proofs between MSR and SCPM.Their formulation imposes conditions on C for equivalence with a given scoring rule S.
- SCPM: SCPM orders specify a limit price π, limit quantity l, and desired-state vector a; the market maker chooses the filled quantity and charge by optimization.The optimization is solved for each arriving order.
- SCPM: SCPM defines state prices as optimal dual variables and charges the trader using the final price and filled order.The optimization also represents worst-case profit through surplus shares and the accumulated-share variable.
- SCPM: Adding the utility term to SCPM enhances the market maker’s risk-taking ability.Without that utility term, the organizer’s objective is interpreted as maximizing worst-case profit.
3. THE UNIFYING FRAMEWORK
The generalized SCPM is a convex optimization framework that adds truthful incremental pricing, efficient cost-function computation, and worst-case loss guarantees. Its utility function determines the market maker’s pricing and risk-related objectives.
- Framework: The generalized SCPM uses a concave utility function in a convex optimization model and retains global optimality, duality, and polynomial computational complexity.The framework extends the original Log-SCPM while preserving its core optimization properties.
- Pricing properties: The SCPM’s pricing function has normalized, nonnegative, consistent, non-decreasing, and integrable properties under the stated utility conditions.Nonnegativity requires a non-decreasing utility function.
- Truthful pricing scheme: Myopically truthful bidding holds for any utility function when traders are charged through the incremental price integral.Truthfulness depends on the charging implementation rather than on the particular utility function.
- Cost function: The incremental pricing integral can be computed through a convex cost function, avoiding explicit integration and reducing the calculation to a single-variable convex optimization problem.The resulting cost function is convex and efficiently computable.
- Worst-case loss: The worst-case loss equals B + C(0) when the market starts with zero shares, and bounded loss requires u(s) − s_i to be bounded above.Worst-case loss computation is itself a convex optimization problem.
4. RELATIONSHIP OF THE SCPM AND THE MSR
The SCPM and market scoring rules are equivalent under a cost-function transformation. The framework also characterizes when SCPM utilities induce proper or strictly proper scoring rules.
- Equivalence: Any proper market scoring rule with cost function C can be represented as an SCPM using the concave utility u(s) = −C(−s).The two models accept the same orders and charge the same prices.
- Properness: Conversely, an SCPM utility induces a proper scoring rule when its derivative spans the probability simplex.This condition is also necessary for properness in the derivation presented.
- Strict properness: Smooth utility functions provide a sufficient condition for strict properness by ensuring a unique optimal price vector.Without uniqueness, traders may have multiple optimal strategies or the market maker may not recover the true belief.
- Examples: The LMSR and Quadratic market scoring rules arise as SCPM special cases with suitable utility functions.Both examples satisfy the stated smoothness-based properness condition.
- Examples: The minimum utility u(s) = min_i s_i yields a proper but not strictly proper scoring rule, whereas a linear utility is not proper.The minimum utility’s sub-gradients span the simplex, while a linear utility has a constant derivative.
5. RISKS FOR THE MARKET MAKER
The SCPM market maker’s utility function can be interpreted as a convex risk measure over revenue outcomes. This dual view connects mechanism choice to the organizer’s willingness to trade expected returns for information about the outcome distribution.
- Risk minimization: The SCPM optimization is equivalent to minimizing a convex risk measure of the market organizer’s outcome-dependent revenue when the utility is non-decreasing.The equivalence preserves the set of optimal accepted-order quantities.
- Risk minimization: Any convex risk measure can potentially define an SCPM market through the utility u(s) = −ρ(Y_s).The constructed utility is concave and increasing.
- Dual interpretation: The dual representation evaluates the worst distribution by trading off reduced expected return against a penalty function.The penalty encodes the organizer’s beliefs and willingness to accept losses while learning the true distribution.
- Dual interpretation: The dual price vector corresponds to the distribution minimizing expected organizer revenue plus the penalty function after accepted orders.Thus, SCPM prices reflect the distribution considered by the organizer in the risk representation.
- Mechanism risk attitudes: The Min-SCPM purely maximizes worst-case return, while LMSR and Log-SCPM can accept orders producing negative returns to support distribution learning.Choosing between LMSR and Log-SCPM depends on whether divergence from a prior or likelihood better represents the organizer’s commitment to learning.
6. DISCUSSION
The discussion shows that the unified framework recasts multiple prediction-market mechanisms through utility-function choices while retaining truthfulness, efficient pricing, properness, and loss analysis. Examples connect exponential and quadratic utilities to LMSR-like and quadratic-scoring-rule markets, with risk attitude determined by the selected utility.
- Framework properties: The framework provides myopically truthful bidding, efficient convex-cost pricing, scoring properness, and worst-case loss guarantees.These properties are presented as general features of the unified convex optimization framework.
- Framework properties: Different mechanisms, including SCPM and scoring-rule markets, differ through their choice of utility function.The paper uses this correspondence to analyze worst-case loss, properness, and risk attitude.
- Exponential-SCPM: The Exponential-SCPM has worst-case loss b log N, is myopically truthful, and induces a strictly proper scoring rule.Its utility is concave, non-decreasing, and separable, and orders can be priced with a convex cost function.
- Quad-SCPM: The Quad-SCPM uses a non-decreasing concave utility so prices remain non-negative and sum to 1, while orders remain myopically truthful and convex-cost priceable.Its loss is bounded by b(∥θ∥2 + 1 − 2 min_i θ_i), and its distance from the prior is measured by the 2-norm.
- Quadratic scoring rule: Replacing the inequality constraint with equality yields the quadratic cost function, while the modified constraint ensures non-negative prices and has a prior-distance interpretation.The resulting market is closely related to the quadratic scoring-rule market.
A.1 Proof of Lemma 3.1
The proof analyzes optimal solutions under two values of a parameter and uses concavity, KKT conditions, and monotonicity to establish the lemma’s claims. Differentiability is not essential because gradients can be replaced by sub-gradients.
- Proof: The proof compares optimal solutions for ε1 and ε2, with ε2 > ε1, and observes that the corresponding optimal x values satisfy x2 > x1.This ordering is obtained from the problem formulation.
- Proof: Concavity supplies an inequality relating utility differences to gradients evaluated at the two optimal solutions.The proof then combines this inequality with KKT-based relations.
- Proof: The KKT conditions establish the relevant price relations, yielding p(ε2)^T a ≥ p(ε1)^T a.This proves the monotonicity claim used in the lemma.
- Proof: The remaining argument invokes integrability of bounded monotone functions on finite intervals.This completes the final claim after the monotonicity observation.
- Proof: The proof extends beyond differentiable utilities by replacing gradients with sub-gradients.The stated differentiability assumption is therefore only for ease of presentation.
B.1 Proof of Lemma 4.3
The proof establishes convexity of the cost function associated with every proper scoring rule and verifies the corresponding risk measure’s structural properties. The arguments rely on properness, monotonicity, concavity, and translation changes of variables.
- Convexity: Properness supplies the inequality needed to establish monotonicity of q1·p(q0 + λq1) as λ increases.The proof compares prices at λ2 and λ1 for λ2 > λ1.
- Convexity: The cost function C(q) is convex in q for every cost function corresponding to a proper scoring rule.The proof reduces convexity to showing q1·p(q0 + λq1) is increasing in λ.
- Cost-function properties: The cost function satisfies the translation relation needed for consistency under adding a constant vector to q.The contradiction argument rules out C(q + d e) > d + C(q).
- Risk measure: The induced risk measure is convex because it minimizes t − u(t e + Z) with concave utility u.Its monotonicity follows from monotonicity of u.
- Risk measure: Translation equivariance follows by substituting t′ = t + α, giving ρ(Z + α) = ρ(Z) − α.The change of variables directly produces the stated relation.
D.1 Properties of the LMSR
The LMSR analysis identifies its worst-case loss and characterizes its risk attitude through a utility-based divergence from a prior. For Hanson’s market, the prior is uniform, while b controls risk tolerance.
- Worst-case loss: The LMSR’s worst-case loss is obtained from C(0) = b log N, with the derivation also verifying B = 0.These quantities are used in the worst-case-loss analysis.
- Risk attitude: The utility formulation measures distance from the prior using Kullback-Leibler divergence LKL(p || θ).The derivation rewrites the optimization in terms of transformed variables before recovering the divergence expression.
- Risk attitude: The prior θ minimizes L(p), while b measures tolerance to risk.In Hanson’s market, θ = e, so the mechanism implicitly assumes a uniform prior.
D.2 Properties of the Log-SCPM
The Log-SCPM has an unbounded worst-case loss, while its parameterization links greater aggregate θ_i to greater risk tolerance.
- The worst-case loss is unbounded because C(0) is unbounded.
- Increasing the aggregate θ_i increases risk tolerance.The passages connect higher aggregate θ_i with increased risk tolerance.
D.3 Properties of the Min-SCPM
The Min-SCPM achieves zero worst-case loss, and its penalty function is represented through a max-min expression.
- The worst-case loss is 0.This follows from C(0) = 0 and B = 0.
- The penalty function L(p) is expressed as a max-min optimization that evaluates the gap between the minimum component of s and p^T s.
- Under the stated quadratic parameterization, the worst-case-loss term B equals −2b.
- A version of the SCPM lacks a convex-risk-minimization representation when its quadratic utility expression is not non-decreasing.
D.5 Properties of the Exponential-SCPM
The Exponential-SCPM has worst-case loss b log N and is strictly proper because its utility gradient spans the simplex continuously.
- The worst-case loss is b log N.
- The mechanism’s risk attitude is obtained by resolving the definition of L(p).
- The mechanism is strictly proper because its utility gradient spans the simplex.The gradient is described as smooth and spanning the simplex.
- The utility function is analyzed on the domain s ≥ −b e, with derivatives and continuity used to establish properness.The derivative is zero at the stated threshold, and the resulting expression spans the simplex continuously.