Signaling Games

Definition and core intuition

A signaling game is a two-stage Bayesian game of incomplete information (see "Bayesian Games", prerequisite). In the canonical form there are two players: a sender (S) who privately knows her type $\theta\in\Theta$ and a receiver (R) who does not. The sender first chooses a signal $s\in S$ — an observable action or attribute — and then the receiver chooses an action $a\in A$. Payoffs depend on the sender's type, the signal, and the receiver's action: $u_S(\theta,s,a)$ and $u_R(\theta,s,a)$. The timing converts private information into an observable signal that can influence posterior beliefs and hence decisions; the Harsanyi transformation (a prerequisite: "Bayesian Games") allows us to treat types as drawn by Nature.

Core example: the Spence model (education as signal). Two worker types, High (H) and Low (L), differ in productivity $\pi_H>\pi_L$. Education level $s\ge0$ is observable but costly: type $t$ pays cost $c_t(s)$. Firms (receivers) observe $s$, form beliefs $\mu(\theta\mid s)$ and pay expected wage equal to expected productivity, e.g. wage schedule $w(s)=\mathbb{E}[\pi_\theta\mid s]$. The worker's payoff is $w(s)-c_\theta(s)$. The signaling question: is there an equilibrium in which different types select different $s$ such that firms infer type from $s$? That is, a separating equilibrium.

Adverse selection vs moral hazard

Signaling arises from adverse selection: one side has private information about a static attribute (quality, cost, productivity). By contrast, moral hazard involves hidden actions taken after contracting (effort choice). Both are central to contracting and mechanism design. In "Mechanism Design", we learned to design allocation and transfer rules to implement desired outcomes subject to incentive compatibility. Signaling is about how agents themselves may influence beliefs; screening is how a principal uses a menu to induce self-selection when facing adverse selection.

Why it matters

Signaling theory explains pervasive phenomena: why schooling credentials exist even when they don't increase productivity, why firms use warranties or prices as signals of quality, and why insurance contracts are differentiated. It provides testable predictions: signals must be more costly for lower types (cost separability), and equilibrium configurations (pooling versus separating) depend on the shape of cost functions and priors.

A very simple numeric intuition

Suppose two types H and L with productivities $\pi_H=10$, $\pi_L=5$. Prior probability of H is $p=0.5$. Education costs are linear: $c_H(s)=s$ and $c_L(s)=3s$. If firms pay $w(s)=\mathbb{E}[\pi\mid s]$, then a separating equilibrium can exist if the high type chooses some $s_H$ and the low type chooses $s_L=0$ so that firms pay $w(s_H)=10$ and $w(0)=\mathbb{E}[\pi\mid s=0]=5$ (believing level 0 comes from L). High's net payoff is $10-s_H$, low's payoff at 0 is $5$. High prefers separation if $10-s_H\ge 5$ i.e. $s_H\le5$. Low won't mimic if $5-3s_H\le5$ i.e. $-3s_H\le0$ which is true for all $s_H\ge0$. So any $s_H\in[0,5]$ could sustain separation under beliefs that assign high type to $s_H$ and low type to 0. This numeric check illustrates the two key constraints you'll formalize: incentive-compatibility (IC) for types and firm beliefs on the path and off-the-path.

Formal equilibrium concept

We use Perfect Bayesian Equilibrium (PBE): strategies for S and R plus beliefs $\mu(\theta\mid s)$ satisfying Bayes' rule on-path and utility maximization given beliefs, and beliefs consistent with Bayes' rule where applicable. Signaling adds the wrinkle that off-path beliefs (how the receiver interprets signals not observed in equilibrium) can make or break equilibria; hence the need for refinements (e.g., the Intuitive Criterion) to rule out implausible beliefs.

This section built directly on "Bayesian Games" (Harsanyi beliefs and BNE) and "Mechanism Design" (incentive constraints will be used heavily).

Goal

Derive algebraic conditions for a separating equilibrium in a typical Spence model. Build the minimal-separation argument and compute explicit thresholds for cost parameters.

Model primitives (formal)

Types: two types, $\theta\in\{H,L\}$ with prior $P(H)=p$, $P(L)=1-p$. Productivities: $\pi_H>\pi_L$. Cost of signal $s\ge0$: $c_\theta(s)$ with $c_H'(s)<c_L'(s)$ for all $s$ (signal is less costly for the high type). Firms are competitive and pay expected productivity: $w(s)=\mathbb{E}[\pi_\theta\mid s]$. Strategies: Sender chooses $s_\theta$, Receiver sets $w(s)=\sum_\theta \pi_\theta \cdot 1\{s\text{ believed from }\theta\}$. In a separating equilibrium suppose $s_H>s_L$ and beliefs assign all mass to H on $s_H$ and to L on $s_L$.

Incentive-compatibility constraints

High type prefers $s_H$ to mimicking $s_L$: $$w(s_H)-c_H(s_H)\ge w(s_L)-c_H(s_L).$$Low type prefers$s_L$to mimicking$s_H$:$$w(s_L)-c_L(s_L)\ge w(s_H)-c_L(s_H).$$Because$w(s_H)=\pi_H$and$w(s_L)=\pi_L$ under the separating beliefs, the ICs become

(High IC) $\pi_H-c_H(s_H)\ge\pi_L-c_H(s_L)\tag{IC_H}$

(Low IC) $\pi_L-c_L(s_L)\ge\pi_H-c_L(s_H)\tag{IC_L}$

Re-arrange. Subtract $\pi_L$ from both sides in (IC_H): $$\pi_H-\pi_L\ge c_H(s_H)-c_H(s_L).$$For (IC_L):$$c_L(s_H)-c_L(s_L)\ge\pi_H-\pi_L.$$So a necessary and sufficient condition for separation with$s_H>s_L$ is

Interpretation: the productivity gap $\Delta$ must lie between the incremental cost of the signal for the high type and that for the low type. If the gap is large, separation is easy; if small, low types can mimic cheaply.

Minimal separating equilibrium (single-crossing, monotone costs)

A standard refinement is the minimal separating equilibrium: choose $s_L=0$ and pick $s_H$ as small as possible to satisfy (IC_L). That is, pick $s_H$ solving

Provided (IC_H) holds automatically (since $c_H$ is cheaper, $c_H(s_H)-c_H(0)\le c_L(s_H)-c_L(0)=\Delta$), this $s_H$ is the smallest education level that deters low-type imitation. This is the Spence-style equilibrium: H gets just enough education to separate.

Concrete numeric example

Let $c_H(s)=s$ and $c_L(s)=3s$. Productivities $\pi_H=10$, $\pi_L=5$, so $\Delta=5$. With $s_L=0$, the minimal $s_H$ solves $c_L(s_H)-c_L(0)=3s_H=5$, so $s_H=5/3\approx1.667$. Check (IC_H): $c_H(s_H)-c_H(0)=1.667\le5$, so holds. Wages: $w(s_H)=10$, $w(0)=5$. Utilities: high gets $10-1.667=8.333$, low gets $5$. Low cannot profitably mimic: if low mimics, his payoff would be $10-3(1.667)=10-5=5$, indifferent; off-path tie-breaking or slight perturbations can deter mimicry strictly.

Proof sketch for uniqueness (monotone pooling vs separation)

If cost functions satisfy single-crossing (i.e. the difference $c_L(s)-c_H(s)$ is increasing in $s$), then any separating equilibrium must be monotone: higher types choose weakly higher signals. Given monotonicity, the minimal separating construction above yields the unique separating equilibrium up to off-equilibrium beliefs and wage ties. The main mathematical reason is that single-crossing implies the lower type's incremental cost of increasing $s$ grows faster than the higher type's, so a single cutoff $s^*$ prevents low-type mimicking.

All formulas in this section had nearby numeric exemplars: $c_H(s)=s$, $c_L(s)=3s$, $\Delta=5$, $s_H=5/3$.

Pooling vs separating equilibria — the mathematics of beliefs

A pooling equilibrium is one in which all types choose the same signal $s^*$. The receiver's posterior is the prior: $\mathbb{E}[\pi\mid s^*]=p\pi_H+(1-p)\pi_L$. Pooling is attractive if the cost of separating exceeds its benefit: sending the same signal saves the higher type the cost of separation while firms pay the pooled expected productivity.

Pooling stability condition (example)

Given $s^*$, types' incentives to deviate determine stability. Suppose $s^*=0$ for both types. If a high-type deviates to $\bar s>0$ and firms interpret $\bar s$ as

a high-type signal and pay $\pi_H$, the deviation payoff is $\pi_H-c_H(\bar s)$; so high will deviate if $\pi_H-c_H(\bar s)>p\pi_H+(1-p)\pi_L-c_H(0)$. There may be no profitable deviation if the cost function makes any separating deviation too costly.

Off-equilibrium beliefs and implausible equilibria

Crucially, pooling equilibria often rely on receivers adopting arbitrary beliefs about off-path signals (signals not used in equilibrium). For example, if firms believe that any $\bar s>0$ was sent only by low types, they will pay $\pi_L$ at $\bar s$, and high types will not deviate. These beliefs can sustain pooling but are often unreasonable: why would the firm believe a costly deviation would be sent by the low type who faces higher costs? Belief-based equilibria are thus fragile.

Equilibrium refinements: the Intuitive Criterion and D1

To eliminate such implausible beliefs we use refinements. The Intuitive Criterion (Cho and Kreps, 1987) rules out beliefs that put probability on types for which deviation is not strictly profitable for any belief-supporting action. D1 (Banks and Sobel) is a stronger refinement comparing types' relative gains from deviation.

Formal statement (Intuitive Criterion)

Let $s'$ be an off-path signal. Consider the set of types $T(s')$ that could potentially gain from deviating to $s'$ against some receiver belief. Formally,

where $s^*$ is the equilibrium signal profile. If some types gain and others don't, the Intuitive Criterion says the receiver should assign zero probability to the types who cannot gain (they wouldn't deviate), and place all mass on those who can. Applying this often rules out pooling equilibria sustained by perverse beliefs.

Concrete example illustrating refinement

Return to the numeric example: $\pi_H=10$, $\pi_L=5$, $p=0.5$, $c_H(s)=s$, $c_L(s)=3s$. Consider pooling at $s=0$ with firms paying $w(0)=7.5$. Suppose there is an off-equilibrium signal $s'=1$. Evaluate gains:

• •High deviating: payoff $10-1=9$ which is greater than $7.5$ (gain of $1.5$).
• •Low deviating: payoff $5-3=2$ which is less than $7.5$ (loss of $5.5$).

Thus only high strictly benefits from deviating. Intuitive Criterion implies that if $s'=1$ is observed, firms should place all mass on H and pay $10$. Consequently pooling at 0 is not stable because H has a profitable deviation and firms' reasonable response makes it profitable. This rules out the pooling equilibrium that relied on the firm believing a deviation would be from L.

Screening mechanisms (principal design vs signaling)

In screening, the uninformed principal (e.g., an employer) offers a menu of contracts to induce self-selection: menu {$(a_1,t_1), (a_2,t_2),...$} where $(a,t)$ are allocation and transfer. The principal solves an optimization subject to incentive compatibility (IC) and individual rationality (IR) constraints — the exact same constraints we used above, but the principal now chooses outcomes rather than passively observing signals. This is exactly "Mechanism Design" territory. For example, an insurer designs deductible-premium pairs such that low-risk choose high deductible/low premium and high-risk choose low deductible/high premium.

Numerical screening sketch

Types H and L with probabilities $p$ and $1-p$. Principal's payoff for type $\theta$ when offering contract $(q,t)$ is $v_\theta(q)-t$ and agent payoff is $t-\phi_\theta(q)$ where $\phi_\theta(q)$ is the agent's cost of providing $q$. The principal maximizes expected payoff subject to IC and IR. For concrete numbers: suppose $v_H(q)=10q$, $v_L(q)=5q$, $\phi_H(q)=q$, $\phi_L(q)=3q$, and $p=0.5$. The principal will choose $q_H>q_L$ and transfers $t_H,t_L$ to satisfy

IC_H: $t_H-\phi_H(q_H)\ge t_L-\phi_H(q_L)$

IC_L: $t_L-\phi_L(q_L)\ge t_H-\phi_L(q_H)$

IR: $t_\theta-\phi_\theta(q_\theta)\ge0$.

One finds, using algebra similar to the Spence ICs, that the principal distorts $q_L$ downward relative to the first-best to reduce informational rent to H, a familiar screening result.

Moral hazard contrast

In moral hazard, types are identical ex ante, but actions are unobserved. Incentive compatibility is dynamic (action choices after contracting), and the principal cannot simply rely on signals; instead she uses performance-contingent payments or monitoring. The mathematics uses incentive constraints across actions rather than types.

All formulas above had concrete numeric checks: the pooling example with $w(0)=7.5$, $s'=1$ gave explicit payoffs $9$ vs $2$; screening sketch used $v_H(q)=10q$, etc.

Price as a signal of quality

A price can be an endogenous costly signal if maintaining a high price is costlier (or less profitable) for low-quality sellers. Consider a monopolist seller of two qualities: high quality (H) with cost $c_H$ and low quality (L) with cost $c_L$, where $c_L<c_H$ — costs are lower for the low-quality seller but consumer valuation differs: buyers value H at $v_H$ and L at $v_L$ with $v_H>v_L$. Suppose buyers infer quality from price $p$: a high price may be interpreted as a signal of high quality, so the high-quality seller may prefer to charge a premium to distinguish herself.

Simple numeric signaling-by-price model

Two sellers types with equal prior $p=0.5$. Buyers' willingness to pay: $v_H=20$, $v_L=10$. Costs for posting a price $p$ are simply marginal costs: $c_H=6$, $c_L=2$. If a seller posts $p$, profit for type $\theta$ is $\pi_\theta(p)=(p-c_\theta)\cdot D(p|beliefs)$ where $D$ is demand conditional on buyer beliefs. If demand is all-or-nothing — buyers buy if they believe seller is H and $p\le v_H$, otherwise don't buy — a separating equilibrium can be: low-type posts $p_L=10$, high-type posts $p_H=20$. Buyers observing $20$ believe H and buy; observing 10 they believe L and buy. Profits: high gets $20-6=14$, low gets $10-2=8$. Check deviations: if low tries to mimic $20$, and buyers believe $20$ is H, low's profit would be $20-2=18$ which is higher than 8, so this naive belief can't hold. Instead, for price to signal credibly, posting a high price must be costlier for L in terms of lost sales or credibility. For instance if demand falls steeply with price for L (reflecting production mismatch or reputational loss), separation can be sustained. The key point is the same IC inequalities: the signal must be relatively costly for low types.

Empirical identification: how to tell signaling from productivity

Applied work often asks: is education a pure productivity-enhancing human-capital investment or mainly a signal? Randomized trials (ability-controlling experiments), natural experiments (compulsory schooling laws), sibling or twin designs, and regression discontinuity can help. In "Mechanism Design" we learned structural identification uses models of incentives to recover primitives; here, one estimates cost functions $c_\theta(s)$ and productivity $\pi_\theta$ to see if observed education choices satisfy the IC inequalities consistent with signaling.

Real-world use cases

• •Labor markets: credentials, credential inflation, degree signaling vs skill acquisition
• •Markets for lemons (Akerlof): adverse selection in used cars; warranties as signals
• •Insurance: deductible and premium menus screen risk types; insurers may also use signals like credit scores
• •Industrial organization: price/quality strategies, warranties, brand building as costly signals
• •Finance: IPO underpricing and costly signals from reputable underwriters

Policy implications

If high education is mostly signaling, public provision of credentials (or standardized testing) can distort incentives and welfare compared to direct subsidies for productivity-enhancing activities. Recognizing the difference matters for designing subsidies, taxes, or regulation of disclosure.

Downstream theoretical connections

Mastery of signaling games enables understanding of: mechanism-design solutions to adverse selection problems (optimal screening, Myerson--Satterthwaite trade-offs in bilateral trade), principal-agent models with hidden actions (moral hazard), dynamic signaling (reputation models, job-market signaling over time), and information design/cheap-talk variants. It also prepares you to read empirical structural papers that estimate unobserved cost distributions and draw welfare conclusions.

Concrete numeric illustration tying elements together

Consider the earlier Spence numbers ($\pi_H=10,\pi_L=5,c_H(s)=s,c_L(s)=3s$). Suppose instead the firm offers a menu: "certificate with wage 10 or no certificate with wage 5". This is a screening design that replicates the separating equilibrium but the firm now commits to wages ex ante. The firm optimally sets the wage spread to extract rents consistent with the ICs we solved earlier; the same numbers produce $s_H\approx1.667$ as the minimal certificate cost for H, matching the signaling model. The only conceptual difference is who commits: in signaling, the sender self-selects; in screening, the principal creates choice.

All formulas and claims above were accompanied by specific numeric exemplars to anchor the algebra in real calculations.