When does a prediction market become a thermodynamic system?

Partition functions, information work, susceptibility, and phase structure in markets for belief

For logarithmic market makers, the thermodynamic correspondence is exact. For real exchanges, it becomes a set of empirical restrictions that data can reject.

Mean-field belief landscapeβJ = 1.18

coupling / inverse temperature

two collective statesm = tanh(βJm)

Type: External literature review
Published: 28 July 2026
Scope: Prediction markets · statistical mechanics
Status: Peer-reviewed foundations; 2026 preprints labeled

An LMSR market is not merely like a thermodynamic system: its cost is a log-partition function, its prices are Gibbs probabilities, and its local market impact is a susceptibility matrix.

Two mechanisms to explore

Move the controls to see how liquidity changes Gibbs prices and how mean-field coupling produces a critical response.

Interactive figure 1

Liquidity as a temperature-like response scale

Hold inventory fixed and change b. Lower values concentrate the Gibbs price; higher values flatten it.

Liquidity parameter b2.20

Outcome 114.5%

Outcome 228.7%

Outcome 356.8%

Prices are computed from p_i = exp(q_i/b) / Σ_j exp(q_j/b) for inventory q = (0, 1.5, 3). The analogy concerns response, not a literal heat bath.

Interactive figure 2

A mean-field market at the critical coupling

Increase βJ through one. The symmetric state loses stability and a collective narrative appears.

Dimensionless coupling βJ0.70

Regimesymmetric

Stable |m|0.000

χ below criticality3.33

This is the Curie–Weiss fixed-point equation m = tanh(βJm + βH) at an infinitesimal positive field. A true nonanalytic phase transition requires the ordered large-system limit.

Prediction markets are often described with the vocabulary of physics: prices “equilibrate,” liquidity “cools,” crowds exhibit “herding,” and correlated beliefs undergo “phase transitions.” Most such language is metaphor. In one important family of automated market makers, however, the correspondence is mathematical: the market maker’s cost is a log-partition function; prices are Gibbs expectations; the liquidity parameter sets a temperature-like response scale; and the Hessian of the cost is a susceptibility matrix.

The distinction matters. An analogy can suggest questions. An isomorphism lets us derive answers.

This note develops the exact result for the logarithmic market scoring rule (LMSR), extends it to coupled binary events and parlay contracts, and then asks which thermodynamic diagnostics survive in limit-order-book markets. The central conclusion is deliberately asymmetric:

An LMSR market over a finite outcome space is an exponential-family thermodynamic system by construction. A conventional prediction exchange is not. It may nevertheless exhibit empirically measurable response, correlation, dissipation, and finite-size scaling that thermodynamics helps organize.

The strongest claims below are identities. The claims about real venues are hypotheses with proposed tests.

1. Cost-function markets as exponential families

Let mutually exclusive outcomes be indexed by \(i=1,\ldots,K\), and let \(q_i\) be the outstanding number of securities paying one unit if outcome \(i\) occurs. A cost-function market maker charges

\Delta C=C(q+\Delta q)-C(q). \tag{1}

Its instantaneous price vector is

p_i(q)=\frac{\partial C}{\partial q_i}. \tag{2}

For the LMSR with liquidity parameter \(b>0\),

C(q)=b\log\sum_{j=1}^{K}\exp(q_j/b). \tag{3}

Define \(Z(q;b)=\sum_j\exp(q_j/b)\). Then

p_i(q)=\frac{\exp(q_i/b)}{Z(q;b)}. \tag{4}

Equations (3)–(4) are already the canonical ensemble. If the microstate is the realized outcome \(i\), and we set an effective energy \(E_i=-q_i\) and inverse temperature \(\beta=1/b\), then

Z=\sum_i e^{-\beta E_i},\qquad p_i=\frac{e^{-\beta E_i}}{Z}. \tag{5}

The sign is economically natural: buying the \(i\)-security raises \(q_i\), lowers its effective energy, and increases its Gibbs weight. The analogy is exact after fixing this dictionary, but the interpretation is not literal. The \(q_i\) are financial inventory coordinates, not molecular energies, and \(b\) is chosen by a market designer rather than set by a heat bath.

The free-energy-like potential is

F(q;b)=-b\log Z(q;b)=-C(q). \tag{6}

The sign reversal reflects the convention that \(C\) is money collected by the market maker. Adding a common constant \(a\) to every inventory coordinate gives \(C(q+a\mathbf 1)=C(q)+a\), while prices are unchanged. This gauge direction expresses the fact that a complete bundle pays one unit in every state.

2. Temperature, response, and susceptibility

Differentiate again:

\frac{\partial p_i}{\partial q_j} =\frac{1}{b}\left(p_i\delta_{ij}-p_ip_j\right). \tag{7}

In matrix form,

\nabla^2 C(q)=\frac{1}{b}\left[\operatorname{diag}(p)-pp^\top\right]. \tag{8}

The bracketed matrix is the covariance of the one-hot outcome vector \(X\), so

b\,\nabla^2 C=\operatorname{Cov}_p(X). \tag{9}

This is a fluctuation–response identity: price response to an infinitesimal inventory perturbation is proportional to equilibrium outcome fluctuation. The matrix is positive semidefinite, has null vector \(\mathbf 1\), and has rank at most \(K-1\). Convexity, probability normalization, and gauge invariance are the same fact viewed three ways.

The parameter \(b\) acts like temperature only in a restricted but precise sense. At fixed \(q\), increasing \(b\) flattens the Gibbs distribution; at fixed price displacement, larger \(b\) requires more order flow. Thus \(b\) controls both entropy and mechanical responsiveness. But thermodynamic temperature is intensive, whereas \(b\) can scale with market size. A useful finite-size convention is \(b=N\tau\), where \(N\) is a liquidity or population scale and \(\tau\) remains intensive.

The entropy of quoted beliefs is

S(p)=-\sum_i p_i\log p_i. \tag{10}

Using \(\log p_i=q_i/b-\log Z\),

C(q)=\sum_i p_iq_i+bS(p). \tag{11}

This is the usual energy–entropy decomposition, translated into inventory coordinates. It also exposes an important qualification: quoted entropy is not the entropy of trader beliefs unless the mechanism has actually aggregated those beliefs without strategic, wealth, or microstructure distortions.

3. Convex duality and the price of information

The log-sum-exp potential is dual to negative Shannon entropy. Up to the gauge direction,

C(q)=\sup_{p\in\Delta_K}\left\{p\cdot q+bS(p)\right\}. \tag{12}

The optimizer is exactly the quoted price in (4). This variational principle is the cleanest bridge among automated market making, maximum entropy, and statistical mechanics.

Suppose a trader changes the price vector from \(p\) to \(p'\) and outcome \(i\) occurs. Because \(q_i=b\log p_i+\lambda\) for a common gauge \(\lambda\), the trader’s realized profit is

\Pi_i=b\log\frac{p_i'}{p_i}. \tag{13}

If the trader’s belief is \(r\in\Delta_K\), expected profit is

\begin{aligned} \mathbb E_r[\Pi] &=b\sum_i r_i\log\frac{p_i'}{p_i}\\ &=bD_{\mathrm{KL}}(r\|p)\\ &\quad-bD_{\mathrm{KL}}(r\|p'). \end{aligned} \tag{14}

An unconstrained risk-neutral trader maximizes this by setting \(p'=r\), obtaining

V_{\mathrm{info}}(r;p)=bD_{\mathrm{KL}}(r\|p). \tag{15}

This is stronger than saying that information “has value.” Under the model, its ideal extractable value is exactly a KL divergence measured in currency units by \(b\). Trading performs a Bregman projection; expected profit is the reduction in informational distance.

The statement depends on assumptions. Budget constraints, risk aversion, transaction fees, position limits, latency, heterogeneous settlement beliefs, and strategic concealment all prevent a trader from moving the public price fully to \(r\). In those cases (15) is a benchmark, not a forecast of realized profit.

The market maker’s worst-case loss is bounded. Starting from \(q=0\), it is at most

L_{\max}=b\log K. \tag{16}

The bound resembles the free-energy cost of selecting one state from \(K\) equiprobable alternatives. It also makes the design tradeoff explicit: larger \(b\) buys smoother, deeper aggregation at a linearly larger subsidy.

4. Many binary events: an Ising market

For \(M\) binary events write a joint outcome as spins \(s=(s_1,\ldots,s_M)\in\{-1,+1\}^M\). A pairwise exponential-family market has

P_\theta(s)=\frac{1}{Z(\theta)} \exp\left(\sum_a h_as_a+\sum_{a<b}J_{ab}s_as_b\right), \tag{17}

with

Z(\theta)=\sum_{s\in\{-1,+1\}^M} \exp\left(\sum_a h_as_a+\sum_{a<b}J_{ab}s_as_b\right). \tag{18}

The \(h_a\) encode marginal pressure and the \(J_{ab}\) encode dependence. Prices of sufficient-statistic securities are expectations:

\frac{\partial\log Z}{\partial h_a}=\langle s_a\rangle,\qquad \frac{\partial\log Z}{\partial J_{ab}}=\langle s_as_b\rangle. \tag{19}

Second derivatives are connected correlations:

\frac{\partial^2\log Z}{\partial h_a\partial h_b} =\langle s_as_b\rangle-\langle s_a\rangle\langle s_b\rangle. \tag{20}

This market quotes a single coherent joint distribution. A marginal security, a conditional security, and a parlay are not priced by separate curves; they are expectations under the same \(P_\theta\). For example, with event indicators \(x_a=(1+s_a)/2\), the price of an all-YES parlay on a set \(A\) is

\pi_A=\mathbb E_\theta\!\left[\prod_{a\in A}x_a\right]. \tag{21}

The computational obstacle is immediate: exact evaluation of \(Z\) generally sums over \(2^M\) states and is intractable for dense graphs. Sparse graphical structure, junction-tree methods, mean-field approximations, Monte Carlo, variational bounds, or restricted securities are not implementation details; they determine which “thermodynamic market” can actually clear.

The 2026 ParlayMarket preprint studies a unified pairwise exponential-family mechanism for base and joint contracts, and APMM develops hierarchical parameter sharing so information from low-leg parlays propagates into higher-leg contracts. These are technically relevant examples, not settled empirical results: both should be labeled preprints, and their loss and convergence guarantees apply under their stated model assumptions.

5. A genuine phase-transition construction

To obtain a phase transition rather than merely borrow the phrase, consider a Curie–Weiss family:

P_N(s)\propto \exp\left[ \beta\left(\frac{J}{2N}\Big(\sum_{a=1}^{N}s_a\Big)^2 +H\sum_{a=1}^{N}s_a\right) \right]. \tag{22}

Let \(m=N^{-1}\sum_as_a\). Standard large-deviation analysis gives a rate-function or free-energy density

\phi(m)= -\frac{\beta J}{2}m^2-\beta Hm +\frac{1+m}{2}\log\frac{1+m}{2} +\frac{1-m}{2}\log\frac{1-m}{2}. \tag{23}

Stationary points satisfy

m=\tanh\!\left[\beta(Jm+H)\right]. \tag{24}

At \(H=0\), the symmetric solution \(m=0\) is unique for \(\beta J\le1\). For \(\beta J>1\), two nonzero solutions appear. Susceptibility in the symmetric phase is

\chi=\left.\frac{\partial m}{\partial H}\right|_{H=0} =\frac{\beta}{1-\beta J}. \tag{25}

The divergence at \(\beta J=1\) is the phase transition. In market language, coupling among related beliefs overwhelms idiosyncratic evidence, and an arbitrarily small field can coordinate the system on one of two macroscopic narratives.

But every finite-\(N\) partition function is a finite sum of exponentials and therefore analytic. There is no literal nonanalytic transition in a finite prediction market. Spontaneous symmetry breaking requires an ordered limit—typically \(N\to\infty\) before \(H\to0\). In observed markets one should look for finite-size scaling, bimodality, hysteresis, and growing susceptibility, not announce a phase transition from a dramatic price chart.

6. Nonequilibrium trading and information work

The equilibrium identities above concern a static LMSR state. A time series of trades is a driven process. Let \(\theta_t\) denote the market state and let \(P_{\theta_t}\) be the instantaneous exponential-family distribution. A protocol that changes \(\theta\) performs financial work, while outcome revelation and trader arrival change the informational environment.

For a small state change, the local cost is

dC=\langle T(X)\rangle_\theta\cdot d\theta, \tag{26}

where \(T(X)\) is the vector of payoff statistics. The quadratic correction is controlled by

g_{ab}(\theta) =\operatorname{Cov}_\theta(T_a,T_b), \tag{27}

the Fisher information metric. This supplies a geometry of market impact: directions with large covariance are highly susceptible, while directions with small covariance require substantial parameter motion to change prices.

In an ideal reversible protocol the market remains near the instantaneous exponential-family state. Fast trading, stale quotes, order batching, latency, strategic splitting, and discrete ticks generate path dependence. A practical excess-work statistic is the realized cost of a path minus the minimum convex-potential difference compatible with its endpoints. Positive excess cost is an entropy-production analogue—but only after fees, spread, and mechanical impact are separately modeled.

A fluctuation–dissipation test can be made operational. Estimate spontaneous short-horizon covariance of price innovations during a stable regime, then estimate the response kernel to plausibly exogenous signed order flow. Under the equilibrium mechanism their normalized forms are linked by (8)–(9). Persistent mismatch, after accounting for asynchronous observation and changing liquidity, is evidence of nonequilibrium dynamics or model misspecification.

7. What survives in real prediction exchanges?

Polymarket documents a hybrid central limit order book with off-chain matching and on-chain settlement. Kalshi exposes binary order books and uses request-for-quote flows for combo markets, each of which has a dedicated book. Neither venue is globally an LMSR canonical ensemble. Their displayed prices can be midpoints, last trades, bids, or asks; liquidity is supplied strategically; and marginal and joint contracts can fragment across books.

Thermodynamic quantities must therefore be estimators, not labels.

Effective liquidity temperature. For a binary contract with log-odds \(\ell=\log[p/(1-p)]\), estimate the local response to signed net order flow \(Q\):

b_{\mathrm{eff}}^{-1} =\frac{\partial \ell}{\partial Q}. \tag{28}

This is state-, horizon-, and side-dependent in a book. Reporting a single number without those qualifiers hides the microstructure.

Susceptibility spectrum. For a vector of related contract prices, estimate the Jacobian of price changes with respect to exogenous order-flow shocks. Its eigenvectors identify collective belief modes; its largest eigenvalue measures amplification. Compare that response matrix with the covariance matrix of unforced fluctuations rather than assuming equality.

Informational value. Use \(b_{\mathrm{eff}}D_{\mathrm{KL}}(r\|p)\) as a counterfactual benchmark for a forecast \(r\), then compare it with feasible profit after spread, depth, fees, and limits. The gap measures implementation friction and strategic constraints.

Entropy production. Construct round-trip or closed-loop trading paths in a stable state. In a conservative cost-function market their line integral vanishes aside from known convex cost. Persistent directional residuals may reveal spread, adverse selection, latency, or inventory management. Calling the residual “entropy production” is useful only if those components are reported.

Criticality. Across comparable market sizes and time-to-resolution bins, test whether susceptibility peaks sharpen, correlation lengths grow across event graphs, distributions become bimodal, and relaxation slows. A power-law fit by itself is not evidence of a critical point.

Lee–Yang-style diagnostics. Zeros of an analytically continued empirical generating function can diagnose sharpening crossovers in finite systems. Because estimated high-order moments are unstable and market samples are nonstationary, such zeros should be accompanied by bootstrap uncertainty and null-model comparisons. They are exploratory diagnostics, not proof of thermodynamic singularities.

8. A falsifiable empirical program

A serious “statistical mechanics of prediction markets” should pre-register observables and failure conditions.

Reconstruct synchronized order books, trades, resolutions, fees, and contract relationships.
Separate mechanical impact from information by using cancellations, quote revisions, and matched-control windows.
Estimate local response tensors and spontaneous covariance on the same clock and state bins.
Test the LMSR fluctuation–response restriction after allowing a time-varying \(b_{\mathrm{eff}}\).
Fit joint exponential families to marginals and parlays, then evaluate held-out calibration and arbitrage consistency.
Search for finite-size scaling only across a defensible family of systems; do not treat time as system size.
Compare against microstructure baselines that contain no thermodynamic criticality.

The thermodynamic model fails usefully if no stable state variable predicts response, if the inferred potential depends strongly on path, if covariance and susceptibility cannot be reconciled even locally, or if a conventional order-book model predicts the same data more parsimoniously. A framework that cannot lose is not physics.

9. Conclusion

The exact insight is compact:

\begin{gathered} \text{LMSR cost}\\[-0.2em] = \text{log partition function},\\[0.35em] \text{prices}\\[-0.2em] = \text{Gibbs expectations},\\[0.35em] \text{impact}\\[-0.2em] = \text{susceptibility}. \end{gathered} \tag{29}

From it follow a bounded-loss mechanism, an entropy-regularized variational principle, and the identity between ideal informed profit and KL-divergence reduction. Coupled event securities lead naturally to Ising-type exponential families, where correlations and parlay prices are derivatives of one joint potential. Mean-field couplings can display a genuine phase transition—but only in a controlled large-system limit.

The broader research opportunity is not to declare markets thermodynamic. It is to measure where their dynamics approximate a potential system, where fluctuation and response separate, and how information, liquidity, and coupling determine collective belief. Thermodynamics is most valuable here when it stops being a metaphor and becomes a collection of restrictions that data can reject.

References

R. Hanson, “Logarithmic Market Scoring Rules for Modular Combinatorial Information Aggregation,” Journal of Prediction Markets 1(1), 3–15 (2007).
J. Abernethy, Y. Chen, and J. Wortman Vaughan, “An Optimization-Based Framework for Automated Market-Making,” EC ’11 (2011).
Y. Chen and D. M. Pennock, “A Utility Framework for Bounded-Loss Market Makers,” UAI 2007; arXiv:1206.5252.
E. T. Jaynes, “Information Theory and Statistical Mechanics,” Physical Review 106, 620–630 (1957).
T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. (Wiley, 2006).
H. Touchette, “The Large Deviation Approach to Statistical Mechanics,” Physics Reports 478, 1–69 (2009).
H.-O. Georgii, Gibbs Measures and Phase Transitions, 2nd ed. (de Gruyter, 2011).
S. Amari and H. Nagaoka, Methods of Information Geometry (AMS/Oxford, 2000).
U. Seifert, “Stochastic Thermodynamics, Fluctuation Theorems and Molecular Machines,” Reports on Progress in Physics 75, 126001 (2012).
G. E. Crooks, “Entropy Production Fluctuation Theorem and the Nonequilibrium Work Relation,” Physical Review E 60, 2721–2726 (1999).
C. Jarzynski, “Nonequilibrium Equality for Free Energy Differences,” Physical Review Letters 78, 2690–2693 (1997).
R. Rana, V. Nadkarni, N. Moshrefi, and P. Viswanath, “ParlayMarket: Automated Market Making for Parlay-style Joint Contracts,” arXiv:2603.22596 (2026), preprint.
N. Moshrefi, R. Rana, and P. Viswanath, “APMM: Automated Parlay Market Maker,” arXiv:2607.18299 (2026), preprint.
Polymarket, “Prices & Orderbook,” platform documentation (accessed July 28, 2026).
Kalshi, “Request for Quote (RFQ)” and “Combos,” platform documentation (accessed July 28, 2026).

Editorial note

“Temperature,” “work,” “entropy production,” and “phase transition” are used literally only where a mathematical definition is supplied. Platform descriptions reflect public documentation accessed on July 28, 2026. The 2026 ParlayMarket and APMM manuscripts are preprints and have not been treated as peer-reviewed consensus.