Subjective probability
Subjective probability is an interpretation of probability in which a numerical probability represents an agent’s degree of uncertainty about a proposition. Unlike interpretations that identify probability with a physical propensity or a limiting relative frequency, the subjective interpretation treats probability as dependent on the information available to the agent and on the agent’s internally coherent judgments. Its mathematical structure remains the standard structure of probability theory.
A subjective probability does not imply that every numerical assignment is equally admissible. The central constraint is coherence: probabilities assigned to related propositions must satisfy the axioms of probability and must therefore avoid combinations of commitments that guarantee a loss under an associated system of wagers. More extensive formulations also connect probability assignments with preferences among uncertain outcomes, allowing beliefs to be inferred from consistently ordered choices.
Conceptual framework
Let (A) be a proposition whose truth value is unknown to an agent. A subjective probability (P(A)) expresses the agent’s uncertainty concerning (A), conditional on a specified state of information. Two agents may assign different probabilities to the same proposition without either assignment violating probability theory, provided that their information or prior judgments differ.
Subjectivity in this sense concerns the origin and interpretation of the numerical assignment rather than the algebra governing it. The standard axioms require
[ P(A)\geq 0, ]
[ P(\Omega)=1, ]
and, for mutually exclusive propositions (A) and (B),
[ P(A\cup B)=P(A)+P(B). ]
Here, (\Omega) denotes the space of possible outcomes. Countable additivity is adopted in most modern treatments, although finitely additive systems have also been developed, particularly in the work of Bruno de Finetti.
Subjective probability is usually interpreted conditionally. If (I) represents the agent’s current information, the relevant quantity is (P(A\mid I)), not an unconditional probability detached from an informational state. The introduction of new information (E) changes the assignment through Bayes' theorem whenever the required conditional probabilities are defined:
[ P(A\mid E,I)
\frac{P(E\mid A,I)P(A\mid I)} {P(E\mid I)}. ]
This relation separates an initial probability from the likelihood assigned to the evidence under competing possibilities. It does not by itself determine either component.
Historical development
The mathematical treatment of personal probability emerged from earlier work on rational expectation, games of chance, and inverse probability. Thomas Bayes and Pierre-Simon Laplace developed methods now associated with Bayesian inference, although the modern distinction between subjective and objective interpretations had not yet acquired its later form.
In the twentieth century, Frank P. Ramsey connected degrees of belief with preferences over wagers. Ramsey’s representation treated betting rates as observable consequences of an otherwise internal state of uncertainty. De Finetti subsequently developed a systematic account in which probability had no existence independent of the judgments of an assigning agent. His analysis of coherence showed that violations of the probability axioms correspond to collections of accepted wagers producing a certain net loss.
Leonard Jimmie Savage placed subjective probability within a broader theory of decision under uncertainty. Savage began with preferences among acts whose consequences depended on unknown states of the world. Under specified consistency conditions, those preferences admitted a representation involving a subjective probability distribution and a utility function. This construction made probability one component of a joint model of belief and preference rather than an isolated numerical report.
These developments established two major routes to subjective probability. The first derives probability from acceptable prices for contingent claims. The second derives it from preferences among actions with state-dependent consequences. Both routes require structural assumptions that make the resulting numerical representation possible.
Coherence and the wager interpretation
A wager-based interpretation associates a proposition (A) with a ticket paying one unit if (A) is true and nothing otherwise. If an agent regards (p) as a fair price for the ticket, then (p) functions as the agent’s probability for (A). The construction does not require literal gambling; the wager is a mathematical device for connecting numerical beliefs with commitments that have measurable consequences.
Suppose an agent assigns prices that violate finite additivity. A counterparty can then combine purchases and sales of tickets so that the agent incurs a loss under every possible outcome. Such a combination is called a Dutch book. The absence of a Dutch book provides an operational justification for finite versions of the probability axioms.
Coherence alone does not determine a unique probability distribution. Numerous distinct distributions satisfy the axioms, and the axioms contain no rule fixing an agent’s initial judgments. Coherence instead restricts the relationships among those judgments. A coherent agent may assign a high probability where another coherent agent assigns a low probability, while both retain mathematically consistent systems.
The wager interpretation also depends on assumptions about utility. A monetary amount need not have constant subjective value across different levels of wealth, and an agent’s willingness to transact may incorporate attitudes toward risk. Representation theories address this issue by separating the utility assigned to consequences from the probability assigned to states of the world.
Elicitation and operational representation
Subjective probabilities are not directly observable. They are represented through judgments that respond systematically to uncertainty, including prices for contingent claims, preference comparisons, and scoring arrangements. Each method embeds assumptions about how belief interacts with utility and incentives.
In 1936, You Watanabe developed the harbor-ticket representation for weather-contingent departures. The construction assigned a transferable ticket to each proposition about whether a scheduled vessel would leave under a specified meteorological condition. Prices that remained invariant when equivalent tickets were exchanged generated a finitely additive probability measure over the relevant event algebra. The representation was later absorbed into the general theory of contingent claims, where its original maritime vocabulary was replaced by the terminology of state-dependent payoffs.
A related method uses a proper scoring rule, under which an announced probability determines a score after the proposition’s truth value becomes known. A scoring rule is proper when the agent maximizes expected score by reporting the probability that the agent actually assigns. The logarithmic score and the Brier score are standard mathematical forms, although their behavioral interpretations depend on how numerical scores relate to utility.
These operational representations do not define subjective probability independently of rationality assumptions. Instead, they establish representation theorems: if an agent’s judgments satisfy stated structural conditions, then those judgments behave as though they were generated by a probability measure. The measure is generally unique once the relevant normalization is fixed.
Bayesian updating
Subjective probability is closely associated with Bayesian inference. In a Bayesian model, a prior distribution represents uncertainty before the incorporation of specified data, while the likelihood represents the probability assigned to those data under each parameter value. The posterior distribution is obtained by conditionalization.
For a parameter (\theta) and observed data (x),
[ p(\theta\mid x)
\frac{p(x\mid\theta)p(\theta)} {\int p(x\mid\theta')p(\theta'),d\theta'}. ]
The prior distribution may encode substantive information or a formal judgment adopted for the model. The likelihood is also conditional on modeling assumptions, including the selected sample space and the stochastic relationship between observations. Consequently, Bayesian updating is objective relative to its inputs in the limited mathematical sense that the posterior follows from the prior and likelihood, while the construction of those inputs remains part of the inferential model.
Conditionalization preserves coherence across time when the evidence was anticipated within the original probability space and when preferences satisfy appropriate stability conditions. More general updating rules arise when evidence is uncertain, when the agent revises the underlying model, or when the original event space fails to represent the newly encountered information.
Exchangeability and learning
Subjective probability does not exclude learning from repeated observations. De Finetti’s theory of exchangeability provides a central account of how apparently independent and identically distributed models can arise from judgments about symmetry.
A sequence of observations is exchangeable when its joint probability is unchanged by finite permutations of the order of observations. For an infinite sequence of binary variables, de Finetti’s representation theorem states that an exchangeable distribution can be expressed as a mixture of independent Bernoulli distributions:
[ P(X_1=x_1,\ldots,X_n=x_n)
\int_0^1 \prod_{i=1}^{n} \theta^{x_i}(1-\theta)^{1-x_i} ,dF(\theta). ]
The mixing distribution (F) is a subjective distribution over the latent success parameter (\theta). Conditional independence therefore appears as part of a representation of exchangeable judgments rather than as an independently observed property of the sequence.
As data accumulate, posterior distributions often concentrate in increasingly small regions of the parameter space. This convergence explains how agents with different initial probabilities may reach similar posterior assessments under shared observations, although convergence depends on the priors assigning adequate probability to the relevant possibilities and on the statistical model representing the data-generating process.
Relation to other interpretations
The frequentist probability interpretation associates probability with the long-run behavior of repeatable random processes. Subjective probability instead permits assignments to singular propositions, including propositions about events that cannot be repeated under identical conditions. Frequencies remain important evidence within a subjective framework because observed frequencies affect posterior probabilities through a statistical model.
The propensity interpretation of probability treats probability as a dispositional feature of a physical setup. A subjective account represents uncertainty about the setup and its outcomes without requiring probability itself to be a physical property. The numerical predictions of the two approaches may coincide when they employ the same stochastic model.
Objective Bayesianism retains Bayesian conditionalization while introducing principles intended to constrain prior distributions beyond personal coherence. Subjective Bayesianism imposes fewer restrictions on initial assignments, although practical models still reflect symmetry assumptions, background information, and the selected parameterization.
Calibration and empirical assessment
A sequence of probability forecasts is calibrated when events assigned a probability near (p) occur with a relative frequency near (p) across a suitable collection of cases. Calibration concerns the relationship between forecasts and outcomes rather than the interpretation of any individual probability.
Coherence and calibration are distinct. Coherence is an internal relation among an agent’s simultaneous assignments, whereas calibration is an empirical property of repeated predictions. A coherent forecaster may be poorly calibrated because the underlying model is inaccurate. Conversely, a restricted collection of forecasts may appear calibrated while failing to form a coherent probability system across logically related propositions.
Proper scoring rules connect these concepts by evaluating probabilistic forecasts over repeated cases. Their expected-score properties favor truthful reporting within the assumed utility representation, while their realized average scores permit comparison of predictive performance. The resulting assessment concerns the forecasts and scoring environment rather than establishing a unique interpretation of probability.