Introduction

The question of how prices form in financial markets has occupied economists, physicists, and mathematicians for over a century. From Bachelier’s pioneering 1900 thesis modeling stock prices as Brownian motion , through the development of rational expectations equilibria by Muth and Lucas , to the modern theory of market microstructure , a central puzzle has persisted: how do markets aggregate the dispersed, incomplete, and often contradictory information held by millions of individual participants into a single number—the price?

In this paper, we argue that the most natural and mathematically precise answer to this question is that markets are distributed Bayesian inference engines. By this we mean something quite specific: the price formation process in a well-functioning market is formally equivalent to a distributed algorithm for computing the posterior distribution over future asset values, given the totality of information available to all market participants.

This perspective is not entirely new. The intellectual roots trace to Friedrich Hayek’s celebrated 1945 essay on the use of knowledge in society , in which he argued that the price system serves as a mechanism for communicating information that no single agent could possess in its entirety. What is new in our treatment is the degree of mathematical precision we bring to this intuition, and the rich connections we draw to adjacent fields: statistical mechanics, information theory, mechanism design, and the theory of complex adaptive systems.

The Central Thesis

Our central claim can be stated in the language of probability theory as follows. Let \(V\) denote the (unknown) fundamental value of an asset, and let \(\mathcal{I}_1, \mathcal{I}_2, \ldots, \mathcal{I}_N\) denote the private information sets of \(N\) market participants. Each agent \(i\) maintains a posterior belief: \[\begin{equation} \label{eq:individual_posterior} \pi_i(V) = \mathbb{P}(V \mid \mathcal{I}_i) \end{equation}\] The market price \(P\) emerges from the interaction of these agents through the trading process. Our thesis is that, under suitable conditions on the market microstructure, the equilibrium price satisfies: \[\begin{equation} \label{eq:market_posterior} P = \mathbb{E}\left[V \,\Big|\, \bigcup_{i=1}^{N} \mathcal{I}_i\right] + \eta \end{equation}\] where \(\eta\) captures noise from liquidity trading, behavioral biases, and microstructural frictions. In other words, the price approximates the conditional expectation of fundamental value given all information available to any participant.

This is a remarkable computational achievement. No central authority collects or processes the information \(\mathcal{I}_i\). No single agent knows the union \(\bigcup_i \mathcal{I}_i\). Yet the decentralized interaction of self-interested traders produces an output that approximates the fully informed Bayesian posterior. The market, in effect, solves a distributed inference problem of enormous complexity.

Outline of the Paper

The paper is organized as follows. Section 2 establishes the Bayesian foundations of belief formation in markets. Section 3 formalizes the information aggregation mechanism through which individual beliefs are synthesized into market prices. Section 4 recasts the Efficient Market Hypothesis as a fixed-point condition on a distributed inference algorithm. Section 5 provides a rigorous treatment of the Hayek information thesis in the language of mechanism design. Section 6 develops the thermodynamic analogy, connecting market dynamics to statistical mechanics. Section 7 analyzes power-law scaling in volatility and returns. Section 8 treats liquidity as an entropic quantity. Section 9 analyzes pure-belief assets (with Bitcoin as the primary case study) as self-referential Bayesian equilibria. Section 10 provides a Bayesian account of bubble formation and collapse. Section 11 develops the analogy between markets and neural networks. Section 12 discusses empirical predictions and tests. Section 13 concludes.

Bayesian Foundations of Market Belief Formation

The Agent as a Bayesian Reasoner

We begin by modeling individual market participants as Bayesian agents. Consider agent \(i\) who maintains beliefs about the fundamental value \(V\) of an asset. At time \(t=0\), the agent holds a prior distribution \(\pi_i^{(0)}(V)\) reflecting her initial assessment before receiving any market-specific information. As information arrives—earnings reports, macroeconomic data, supply chain signals, geopolitical developments—the agent updates according to Bayes’ rule.

This recursive updating has a well-known product form. Defining the information set \(\mathcal{I}_i^{(T)} = \{s_i^{(1)}, \ldots, s_i^{(T)}\}\) and assuming conditional independence of signals given \(V\), the posterior takes the form: \[\begin{equation} \label{eq:product_form} \pi_i^{(T)}(V) \propto \pi_i^{(0)}(V) \cdot \prod_{t=1}^{T} f(s_i^{(t)} \mid V) \end{equation}\]

This factorization is crucial because it means each signal contributes a multiplicative “likelihood ratio” to the posterior, and the order in which signals arrive is irrelevant to the final belief—a property known as exchangeability of Bayesian evidence.

Heterogeneity of Information and Priors

A critical feature of real markets is that agents are informationally heterogeneous. This heterogeneity takes two forms:

1. Signal heterogeneity: Different agents observe different signals. A semiconductor analyst observes chip production data; a macroeconomist observes interest rate signals; a supply chain manager observes shipping volumes. Formally, the signal spaces \(\mathcal{S}_i\) may differ across agents.

2. Prior heterogeneity: Even agents observing identical signals may interpret them differently due to different priors \(\pi_i^{(0)}\). A value investor and a momentum trader may assign different prior probabilities to the same fundamental scenarios.

Let us formalize the aggregate information structure. Define the economy-wide information set as: \[\begin{equation} \mathcal{I}^* = \bigcup_{i=1}^{N} \mathcal{I}_i \end{equation}\] and the fully informed posterior as: \[\begin{equation} \label{eq:full_posterior} \pi^*(V) = \mathbb{P}(V \mid \mathcal{I}^*) \end{equation}\]

The fundamental question of information aggregation is: under what conditions does the market price \(P\) reflect \(\pi^*(V)\)?

The Log-Likelihood Representation

For analytical tractability, it is often useful to work in log-space. Define the log-belief of agent \(i\): \[\begin{equation} \ell_i(V) = \log \pi_i(V) \end{equation}\]

Under the product form [eq:product_form], the log-posterior decomposes additively: \[\begin{equation} \label{eq:log_decomposition} \ell_i^{(T)}(V) = \ell_i^{(0)}(V) + \sum_{t=1}^{T} \log f(s_i^{(t)} \mid V) + \text{const.} \end{equation}\]

This additive structure in log-space is deeply analogous to energy decomposition in statistical mechanics, a connection we exploit extensively in Section 6.

The Information Aggregation Mechanism

Prices as Sufficient Statistics

The notion that prices aggregate information can be made precise using the concept of sufficient statistics from mathematical statistics.

Price sufficiency is a strong condition. It states that knowing the price alone conveys exactly as much information about \(V\) as knowing the entirety of all agents’ private information. Under what conditions is this achievable?

This result provides the mathematical foundation for the intuition that prices “compress” information. In the Gaussian model, an \(N\)-dimensional vector of private signals is losslessly compressed into a single price, with the weights determined by signal precision. Agents with more precise information (lower \(\sigma_i^2\)) receive greater weight in the consensus—a formal vindication of the “smart money” concept.

The Grossman–Stiglitz Paradox and Noisy Rational Expectations

A fundamental tension arises in the information aggregation story. If prices perfectly reveal all information, then no agent has an incentive to acquire costly information—the Grossman–Stiglitz paradox . In equilibrium, there must be enough noise in prices to compensate informed traders for their information acquisition costs.

The resolution lies in the noisy rational expectations equilibrium (NREE) framework. Let \(P\) denote the market price and \(z\) denote the aggregate noise trading demand (exogenous liquidity shocks). In the NREE, the price takes the form: \[\begin{equation} \label{eq:nree} P = \alpha_0 + \sum_{i=1}^{N} \alpha_i s_i + \beta z \end{equation}\] where \(\alpha_i\) are endogenously determined coefficients reflecting each signal’s informational weight, and \(\beta z\) is the noise component. The price is partially revealing: observing \(P\) allows uninformed traders to extract a noisy estimate of \(\sum_i \alpha_i s_i\), but the presence of \(z\) prevents perfect extraction.

This establishes that the market’s inference is asymptotically perfect as noise diminishes, but in practice always leaves a residual informational gap that sustains the incentive for information production.

Message Passing on Information Networks

The process by which individual beliefs are aggregated into market prices can be understood through the lens of message-passing algorithms on graphical models. Consider the market as a factor graph where:

• Variable nodes represent the unknown value \(V\) and each agent’s information \(\mathcal{I}_i\).

• Factor nodes represent the trading interactions between agents.

• Messages are the orders (bids and asks) that agents submit, encoding their beliefs.

The market clearing condition requires that buy and sell pressures balance: \[\begin{equation} \label{eq:market_clearing} \sum_{i=1}^{N} D_i(P; \mathcal{I}_i) = 0 \end{equation}\] where \(D_i(P; \mathcal{I}_i)\) is agent \(i\)’s excess demand at price \(P\). The solution \(P^*\) to this equation is the market-clearing price, which implicitly aggregates all private information through the demand functions.

This interpretation reveals the market as running a form of belief propagation—each agent “broadcasts” her beliefs through her demand function, the market-clearing mechanism “combines” these messages, and the resulting price is the “consensus belief.” This connection to belief propagation on graphical models provides a principled framework for understanding convergence properties, approximation quality, and failure modes (such as loopy belief propagation leading to non-convergence, analogous to market instability).

The Efficient Market Hypothesis as a Fixed-Point Condition

Classical Formulations

The Efficient Market Hypothesis, in its various forms, asserts that market prices reflect available information. We now show that each form corresponds to a specific fixed-point condition on the distributed inference algorithm.

Convergence of the Inference Algorithm

The natural question is whether the distributed inference algorithm converges to the fixed point. We analyze this in the framework of iterated expectations.

Consider a discrete-time market where at each period \(t\), agent \(i\) updates her belief based on (a) her private signal \(s_i^{(t)}\), and (b) the previous market price \(P_{t-1}\) (which she rationally interprets as containing information from other agents). Define the belief-update operator: \[\begin{equation} \label{eq:belief_operator} T_i[\pi](V) = \frac{f(s_i^{(t)} \mid V) \cdot g(P_{t-1} \mid V) \cdot \pi(V)}{\int f(s_i^{(t)} \mid v) \cdot g(P_{t-1} \mid v) \cdot \pi(v) \, dv} \end{equation}\] where \(g(P \mid V)\) is the agent’s model of price formation.

The proof follows from results in social learning theory, specifically the theorem that Bayesian agents who observe each other’s actions will eventually reach consensus on the posterior distribution, provided the signal structure is sufficiently rich . The market price serves as a public signal that facilitates this convergence.

The Hayek Information Thesis: A Mechanism Design Perspective

Hayek’s Original Insight

Hayek’s seminal argument can be restated in modern information-theoretic language. Consider an economy with \(N\) agents, each possessing local information \(\mathcal{I}_i\) about the relevant economic variables. The total information content of the economy is: \[\begin{equation} H(\mathcal{I}^*) = H\left(\bigcup_{i=1}^{N} \mathcal{I}_i\right) \end{equation}\] where \(H(\cdot)\) denotes Shannon entropy. In general, this quantity is enormous—the “knowledge of the particular circumstances of time and place” that Hayek emphasized.

Hayek’s insight is that the price system achieves a dramatic compression: \[\begin{equation} \label{eq:hayek_compression} H(P) \ll H(\mathcal{I}^*) \quad \text{yet} \quad I(V; P) \approx I(V; \mathcal{I}^*) \end{equation}\] where \(I(X;Y)\) denotes mutual information. The price \(P\) has vastly lower entropy (information content) than the aggregate information set, yet it captures nearly all the information relevant to the fundamental value \(V\).

Prices as Optimal Compressed Messages

We can formalize the compression property using rate-distortion theory. The “rate” is the entropy of the price signal \(H(P)\), and the “distortion” is the information loss \(I(V; \mathcal{I}^*) - I(V; P)\).

This result shows that the market price is an optimally compressed representation of the aggregate information—it achieves the information-theoretic limit on how much information about \(V\) can be conveyed by a single real number.

Mechanism Design Foundations

The price mechanism can be understood as a solution to a mechanism design problem. The social planner seeks to aggregate information from strategic agents who may have incentives to misreport their beliefs. The celebrated revelation principle guarantees that any outcome achievable by any mechanism can be achieved by an incentive-compatible direct mechanism.

This provides the mechanism-design foundation for why markets succeed as inference engines: the incentive structure of competitive markets naturally elicits truthful revelation of beliefs through the medium of trading activity. A trader who believes a stock is undervalued has a direct financial incentive to buy, thereby pushing the price toward her assessment. The alignment of private incentives with social information aggregation is the mechanism-design miracle underlying market efficiency.

The Thermodynamic Structure of Financial Markets

Markets and Statistical Mechanics: The Core Analogy

We now develop what is perhaps the deepest structural insight of this paper: the formal analogy between market dynamics and the statistical mechanics of many-body systems. This connection goes beyond mere metaphor—the mathematical structures are isomorphic.

Consider the log-posterior decomposition from Equation [eq:log_decomposition]. Define the market energy function: \[\begin{equation} \label{eq:market_energy} E(V) = -\sum_{i=1}^{N} \ell_i(V) = -\sum_{i=1}^{N} \log \pi_i(V) \end{equation}\]

The aggregate market belief then takes the Boltzmann form: \[\begin{equation} \label{eq:boltzmann} \pi_{\text{market}}(V) \propto e^{-E(V)/T} \end{equation}\] where we have introduced a market temperature parameter \(T\) that controls the sharpness of the distribution.

Free Energy and Market Equilibrium

In statistical mechanics, equilibrium is characterized by the minimization of free energy \(F = E - T \cdot S\), where \(S\) is entropy. The market analogue is:

This result reveals a deep structural principle: the market equilibrium is determined by a trade-off between fitting the data (minimizing energy, i.e., being consistent with all agents’ information) and maintaining appropriate uncertainty (maximizing entropy). This is exactly the maximum-entropy principle of Jaynes , applied to the market context.

The Partition Function and Market Observables

In statistical mechanics, the partition function \(Z\) encodes all thermodynamic information about a system. The market analogue is:

\[\begin{equation} \label{eq:partition} Z(T) = \int e^{-E(V)/T} \, dV \end{equation}\]

From the partition function, we can derive all macroscopic market observables:

\[\begin{align} \text{Price (expected value):} \quad P &= -T \frac{\partial \log Z}{\partial \mu} \label{eq:price_from_Z} \\ \text{Volatility (variance):} \quad \sigma^2 &= T^2 \frac{\partial^2 \log Z}{\partial \mu^2} \label{eq:vol_from_Z} \\ \text{Market entropy:} \quad S &= \frac{\partial (T \log Z)}{\partial T} \label{eq:entropy_from_Z} \end{align}\]

where \(\mu\) parameterizes shifts in the energy function. These relations are the market analogues of the thermodynamic relations connecting free energy to internal energy, heat capacity, and entropy.

Phase Transitions in Market Beliefs

Perhaps the most dramatic consequence of the thermodynamic framework is the possibility of phase transitions—qualitative changes in the market’s collective belief structure as parameters are varied continuously.

This provides a thermodynamic explanation for market crashes: they are first-order phase transitions in the collective belief landscape. The market abruptly transitions from one basin of attraction (e.g., “the company is viable”) to another (“the company will fail”), with the price exhibiting a discontinuous jump. The crash is not a breakdown of the Bayesian mechanism—it is the Bayesian mechanism operating correctly in the presence of a bimodal posterior.

Power-Law Scaling in Financial Markets

Empirical Evidence for Fat Tails

One of the most robust empirical findings in financial economics is that asset returns exhibit fat tails—the probability of extreme price movements is much larger than predicted by Gaussian models. Specifically, the probability of a return \(r\) exceeding a threshold \(x\) follows a power law: \[\begin{equation} \label{eq:power_law} \mathbb{P}(|r| > x) \sim x^{-\alpha} \end{equation}\] with the tail exponent \(\alpha\) typically in the range \(2 < \alpha < 5\) for equity markets, with \(\alpha \approx 3\) being a commonly observed value—the so-called inverse cubic law .

The Criticality Hypothesis

The power-law scaling of returns can be understood through the thermodynamic framework developed in Section 6. In statistical mechanics, power-law distributions emerge naturally at critical points—phase transitions where the system exhibits scale-invariance.

The mechanism for self-organized criticality in markets can be understood as follows. Consider the market temperature \(T\) as defined in Definition 13. At high temperatures (high uncertainty), agents trade conservatively and price movements are small. At low temperatures (strong consensus), agents trade aggressively and prices are stable near the consensus value. The critical regime occurs at intermediate temperatures where neither effect dominates, and the system exhibits maximal sensitivity to perturbations.

Renormalization Group for Market Dynamics

The renormalization group (RG) provides a framework for understanding the universality of critical exponents. In the market context, the RG transformation corresponds to coarse-graining the time scale of observation.

At the critical point, the return distribution is a fixed point of the RG transformation (up to rescaling), which implies the scale-invariance and power-law behavior observed empirically. The tail exponent \(\alpha\) is determined by the eigenvalues of the linearized RG transformation around the fixed point: \[\begin{equation} \alpha = d + \frac{\log \lambda_1}{\log b} \end{equation}\] where \(d\) is the “dimension” of the market (related to the effective number of independent information sources) and \(\lambda_1\) is the largest relevant eigenvalue.

Volatility Clustering and Long Memory

Another robust empirical finding is volatility clustering: large price movements tend to be followed by large movements (of either sign), and small movements tend to follow small movements. The autocorrelation of absolute returns decays slowly: \[\begin{equation} \label{eq:vol_clustering} \text{Corr}(|r_t|, |r_{t+\tau}|) \sim \tau^{-\beta} \end{equation}\] with \(\beta \approx 0.3\), indicating long memory in the volatility process.

In our thermodynamic framework, this long memory arises from the slow relaxation dynamics near the critical point. The market’s “memory” of past volatility regimes reflects the critical slowing-down phenomenon: perturbations to the system decay as power laws rather than exponentials, because the correlation length diverges at criticality.

Liquidity as an Entropic Measure

The Information-Theoretic Nature of Liquidity

Liquidity—the ease with which an asset can be traded without significantly impacting its price—is one of the most important yet poorly understood concepts in financial economics. We propose that liquidity is fundamentally an entropic quantity, measuring the degree of “disorder” or “flexibility” in the market’s belief structure.

The intuition is as follows. When there is a wide diversity of beliefs about an asset’s value (high entropy), the market has many potential counterparties for any given trade—some agents want to buy and others want to sell at any given price. This diversity of opinion is the source of liquidity. Conversely, when all agents agree on the value (low entropy), there are no natural counterparties, and the market is illiquid.

The Liquidity–Volatility Relationship

The entropic framework provides a natural explanation for the well-documented inverse relationship between liquidity and volatility:

Liquidity Crises as Entropy Collapse

A liquidity crisis occurs when market participants simultaneously converge to similar beliefs (typically pessimistic), causing entropy to collapse and liquidity to evaporate. In thermodynamic terms, this is a condensation phenomenon: the system transitions from a high-entropy “gaseous” phase (diverse beliefs, high liquidity) to a low-entropy “condensed” phase (homogeneous beliefs, illiquidity).

The entropy framework also explains the paradox of market making: market makers provide liquidity precisely because they maintain heterogeneous beliefs relative to the marginal trader. A market maker who agreed with the consensus would have no reason to offer liquidity. The market maker’s “edge” is, in information-theoretic terms, the divergence between her belief and the market’s consensus belief.

Pure-Belief Assets: Bitcoin as a Self-Referential Bayesian Equilibrium

Assets Without Fundamentals

Traditional asset pricing theory relates prices to discounted expected cashflows: \[\begin{equation} \label{eq:dcf} P = \sum_{t=1}^{\infty} \frac{\mathbb{E}[C_t]}{(1+r)^t} \end{equation}\] where \(C_t\) is the cashflow at time \(t\) and \(r\) is the discount rate. For assets like Bitcoin, which produce no cashflows (\(C_t = 0\) for all \(t\)), the traditional framework assigns a value of zero.

Yet Bitcoin’s market capitalization has reached over one trillion dollars. How can our Bayesian framework account for this?

Self-Referential Value

The resolution lies in the observation that a pure-belief asset’s value is self-referential: the fundamental value depends on the belief that the asset has value. Formally:

For Bitcoin, the function \(\Phi\) encodes the relationship between adoption probability and value: \[\begin{equation} \label{eq:bitcoin_phi} \Phi(p) = p \cdot V_{\text{max}} \cdot g(\text{network size}) \cdot h(\text{scarcity}) \end{equation}\] where \(V_{\text{max}}\) is the theoretical maximum value if Bitcoin achieves full adoption as a global store of value, \(g\) captures network effects (Metcalfe’s law), and \(h\) captures the scarcity premium from the fixed supply of 21 million coins.

Multiple Equilibria and Coordination

The self-referential structure [eq:self_ref] typically admits multiple equilibria:

This multiple-equilibrium structure explains the volatility of pure-belief assets: small perturbations in collective belief can shift the system between equilibria, causing dramatic price swings. It also explains the role of narratives in sustaining value: the non-trivial equilibrium is sustained by a self-reinforcing story (“Bitcoin is digital gold”) that coordinates beliefs.

The Market Capitalization Illusion

A critical insight from the Bayesian framework concerns the relationship between market capitalization and actual wealth.

This explains why markets can “lose trillions” overnight: the market capitalization was never actual wealth—it was a belief-weighted extrapolation of the marginal trade price to the total supply. When beliefs shift, the extrapolated number changes dramatically, even though the actual flow of funds is small.

A Bayesian Theory of Bubbles

Bubbles as Bayesian Overweighting

In the framework developed above, a bubble occurs when the market posterior systematically overestimates the probability of favorable outcomes. We formalize this as follows.

Mechanisms for Bubble Formation

Several mechanisms within the Bayesian framework can generate bubbles:

Correlated Signals and Common Prior Misspecification

If agents receive signals that are positively correlated (e.g., they all read the same analyst reports), the market may double-count information:

Bayesian Herding

In sequential markets, information cascades can form when agents rationally ignore their private signals in favor of the public signal embedded in the price:

Bubble Collapse as Phase Transition

In the thermodynamic framework, bubble collapse corresponds to a first-order phase transition as described in Theorem 18. The bubble regime corresponds to the metastable state of the double-well energy function: the market sits in a local minimum of the free energy (the “bubble valuation”) that is not the global minimum (the “fundamental valuation”). The transition from the metastable to the stable state—the crash—is triggered by a sufficiently large perturbation that pushes the system over the energy barrier between the two states.

The Market as a Neural Network

Structural Correspondence

The analogy between financial markets and neural networks is more than superficial. We now make it precise.

The Hopfield Market Model

The connection can be made even more precise using the Hopfield network framework . Consider a market with \(N\) agents, each holding a position \(\sigma_i \in \{-1, +1\}\) (short or long). The “energy” of a market configuration \(\boldsymbol{\sigma} = (\sigma_1, \ldots, \sigma_N)\) is: \[\begin{equation} \label{eq:hopfield} E(\boldsymbol{\sigma}) = -\frac{1}{2} \sum_{i \neq j} J_{ij} \sigma_i \sigma_j - \sum_{i} h_i \sigma_i \end{equation}\] where \(J_{ij}\) represents the interaction between agents \(i\) and \(j\) (e.g., herding tendencies, contrarian behavior), and \(h_i\) is agent \(i\)’s private signal (bias toward long or short).

This provides a formal mechanism for the observation that markets often “remember” key price levels—these are the stored patterns in the Hopfield network, and the market dynamics naturally converge to these attractor states.

Deep Learning and the Hierarchical Market

Modern financial markets exhibit a hierarchical structure that mirrors deep neural networks:

1. Input layer: Raw data (earnings, economic indicators, news feeds).

2. First hidden layer: Individual analysts and traders who process raw data into assessments.

3. Second hidden layer: Hedge funds and institutional investors who aggregate analyst assessments.

4. Third hidden layer: Market makers who aggregate institutional order flow.

5. Output layer: The market price.

Each layer performs a nonlinear transformation of its inputs, progressively extracting more abstract features—precisely the operation of a deep neural network. The “training signal” is profit and loss: agents that systematically misprice assets lose capital and exit the market (analogous to pruning ineffective neurons), while successful agents accumulate capital and gain influence (analogous to strengthening connections).

Empirical Predictions and Testable Implications

The theoretical framework developed in this paper generates several novel, testable predictions that distinguish it from alternative theories of market behavior.

Prediction 1: Precision-Weighted Information Aggregation

This prediction can be tested by comparing the speed of price adjustment following information releases of varying precision: earnings announcements (high precision) vs. analyst opinion changes (moderate precision) vs. social media sentiment shifts (low precision).

Prediction 2: Entropy–Liquidity Correlation

Prediction 3: Critical Scaling Near Regime Changes

Prediction 4: Market Temperature and Monetary Policy

Prediction 5: Information Content of Order Flow

Conclusion

We have developed a comprehensive theoretical framework characterizing financial markets as distributed Bayesian inference engines. This perspective unifies and extends several major threads in economic and financial theory:

1. Hayek’s information aggregation thesis receives a rigorous information-theoretic formalization, in which prices are shown to be optimally compressed representations of dispersed private information, achieving the rate-distortion bound.

2. The Efficient Market Hypothesis is recast as a fixed-point condition on a distributed inference algorithm, with convergence guaranteed under mild conditions on the information structure and market microstructure.

3. Market thermodynamics provides a unifying mathematical framework in which price formation corresponds to free energy minimization, market crashes to first-order phase transitions, and power-law volatility to self-organized criticality.

4. Liquidity receives a natural information-theoretic interpretation as the entropy of the market’s belief distribution, explaining the inverse liquidity–volatility relationship and the dynamics of liquidity crises.

5. Pure-belief assets such as Bitcoin are understood as self-referential Bayesian equilibria, with the multiple-equilibrium structure explaining both their existence and their extraordinary volatility.

6. Bubbles and crashes are formalized as systematic Bayesian estimation errors (correlated signals, information cascades) and their resolution through first-order phase transitions.

7. The market–neural network analogy is made precise, with formal correspondences between market components and neural network architecture.

The framework generates novel empirical predictions regarding the speed of information incorporation, the relationship between belief diversity and liquidity, pre-crash criticality signatures, and the information content of order flow. These predictions provide a roadmap for empirical testing and further theoretical development.

Perhaps the deepest implication of this work is philosophical. If markets are indeed distributed inference engines operating near criticality, then they are best understood not as passive reflectors of “value” but as active computational devices that discover probabilistic truths about the future through the decentralized interaction of self-interested agents. The price is not a number—it is a prediction, a probability, a compressed representation of collective intelligence about an uncertain future. Understanding markets in this way—as machines for discovering reality under uncertainty—opens new avenues for both theoretical analysis and practical design of more efficient and robust financial systems.

GrokRxiv DOI: 10.48550/GrokRxiv.2603.04.markets-bayesian-inference 4 March 2026