In traditional economics, a market is often viewed as a simple clearinghouse where supply meets demand. However, a growing body of research suggests a more profound interpretation: markets are decentralized computers designed to solve complex optimization problems. By leveraging the mathematical framework of convex duality, researchers have shown that the mechanisms used to elicit information, allocate resources, and manage financial risk are often different mathematical representations of the same underlying process.
The Symmetry of Information and Risk
The foundation of this perspective lies in the equivalence between how we reward accuracy (scoring rules) and how we manage market liquidity (cost functions). In prediction markets, a market maker must set prices that aggregate the private information of traders into a collective forecast. Agrawal et al. et al. established a "Grand Unified Theory" for this process, proving that proper scoring rules and sequential convex pari-mutuel mechanisms are mathematically equivalent. This means that any strategy for eliciting information can be reformulated as a resource allocation problem.
This duality extends to risk management. Agrawal et al. et al. further demonstrated that proper scoring rules share a dual relationship with convex risk measures. This allows market designers to calculate safety bounds, such as worst-case loss, by treating the market as a minimization problem of a specific risk measure. Similarly, Chen et al. et al. proved that Market Scoring Rules and cost-function-based markets—like those introduced in the Yoopick system by Goel et al. et al.—are simply different ways to represent the same information aggregation process.
Markets as Distributed Optimization Algorithms
If a market is an optimization engine, then the act of trading is the algorithm that runs it. Hu et al. et al. showed that prediction markets, where agents trade to minimize their own individual risk, collectively function as a distributed optimization system for global machine learning objectives. Using Fenchel’s duality, the authors linked the collective minimization of agent risks to the dual of standard problems like logistic regression.
The dynamics of how these markets reach equilibrium also mirror well-known optimization techniques:
- Randomized Subspace Descent: Frongillo et al. et al. proved that individual trading actions function as steps in a global stochastic optimization process, guaranteeing that the market will converge to an equilibrium that minimizes total market risk.
- Mirror Descent: In Fisher markets—where buyers have fixed budgets—Cole et al. et al. showed that natural "Proportional Response" updating by buyers is equivalent to the Mirror Descent algorithm, leading to fast, linear convergence rates.
Unifying Economic Models through Duality
Convex duality also provides a "Rosetta Stone" for unifying disparate economic models. Historically, the Eisenberg-Gale and Shmyrev programs were seen as separate approaches to finding market equilibria. However, Cole et al. et al. demonstrated that these two programs are actually duals of each other under a simple change of variables.
This unification allows researchers to apply tools from one domain to solve problems in another. For instance, Végh showed that the Eisenberg-Gale program is a specific instance of a concave generalized network flow, enabling combinatorial algorithms to solve market equilibrium problems. More recently, Kroer et al. and Dominik Peters applied "double duality" to the Nash social welfare objective to solve the Lindahl equilibrium for public goods, creating a virtual market that remains stable even under strict funding caps.
From Theory to Decentralized Finance
The insights gained from this duality are now being applied to modern financial infrastructure. The Automated Market Makers (AMMs) that power decentralized finance (DeFi) protocols, such as those described by Capponi et al. et al., rely on constant product formulas to maintain liquidity.
Frongillo et al. et al. (2023) provided an axiomatic characterization of these Constant Function Market Makers (CFMMs), proving they are mathematically equivalent to the prediction markets used for forecasting. This connection, also explored by Abernethy et al. et al., shows that the pricing mechanisms used to aggregate crowd beliefs are functionally identical to the loss functions used in machine learning and the liquidity pools used in DeFi. As Denizalp Goktas argues, this unified framework allows us to treat market price discovery as a robust min-max optimization problem, paving the way for more efficient and decentralized marketplaces.
Key Foundational Work
A Unified Framework for Dynamic Pari-Mutuel Information Market Design (2009) — contributed 2 concept elements to this direction.
Yoopick: A Combinatorial Sports Prediction Market (2008) — contributed 1 concept elements to this direction.
Combinatorial Prediction Markets (2008) — contributed 1 concept elements to this direction.