Glosten–Milgrom Order-Flow Component

• The Glosten–Milgrom order-flow component is the information-driven adjustment in prices reflecting adverse selection from informed trading.
• It arises as dealers update conditional expectations based on observed buy and sell orders, generating spreads in both single- and multi-asset frameworks.
• Modern formulations extend this concept to limit-order-book dynamics, linking unexpected order flow to the permanent and transient impacts on price movements.

The Glosten–Milgrom order-flow component is the information, or adverse-selection, component of prices and spreads that arises when liquidity providers infer asset value from observed buy and sell orders. In the canonical Glosten–Milgrom setting, a competitive, risk-neutral dealer sets transaction prices equal to conditional expectations of value given order direction, so a buy shifts price upward and a sell shifts it downward because order flow is informative about private information. In that benchmark formulation, the bid–ask spread is generated entirely by the informational content of order flow rather than by inventory or order-processing costs (Gerig et al., 2010, Touzo et al., 2020).

1. Classical definition in the one-asset Glosten–Milgrom model

In the baseline formulation, there is a single asset with terminal value $\tilde V$, informed traders know $\tilde V$, uninformed traders buy and sell with equal probability for exogenous reasons, and a competitive risk-neutral dealer posts quotes equal to conditional expectations of value given the incoming order sign. The core pricing rule is therefore

$$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$

so the spread is the difference between the value inferred from a buy and the value inferred from a sell (Gerig et al., 2010).

A particularly transparent binary-value specification takes

$\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$

with $\mathbb{E}[\tilde V_i]=p_i$, and lets a fraction $\gamma_i\in(0,1)$ of orders be informed. In the single-security environment, the competitive liquidity provider sets

$$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$

so the ask is $p_i+\gamma_i r_i$, the bid is $p_i-\gamma_i r_i$, and the spread is

$\Delta_i = 2\gamma_i r_i.$

This makes the order-flow component explicit: it scales with the probability of informed trading $\tilde V$0 and with the size of the value jump $\tilde V$1 (Gerig et al., 2010).

In the Bernoulli formulation with $\tilde V$2, the same logic appears as

$\tilde V$3

where $\tilde V$4 is the observed order history. The realised transaction price equals the ask after a buy and the bid after a sell, while the ex ante mid-price is $\tilde V$5. The price process is a martingale, so order flow is the sole driver of belief revisions in this model (Touzo et al., 2020).

The essential implication is that the “order-flow component” is not an auxiliary adjustment layered onto prices. In the benchmark model it is the whole spread. The only reason a buy and a sell receive different prices is that they carry different information about the latent value (Gerig et al., 2010).

2. Bayesian filtering, sufficient statistics, and endogenous information

The informational mechanism is a sequential filtering problem. In the discrete-time Bernoulli model, conditional on the true state $\tilde V$6, the order $\tilde V$7 satisfies

$\tilde V$8

where $\tilde V$9 is the fraction of informed traders. Bayes’ rule then yields a posterior recursion for $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$0, so each order updates the market maker’s belief by a likelihood-ratio step. Because conditional on $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$1 the orders are i.i.d., the order history is summarized by the cumulative number of buys $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$2, and the explicit ask and bid can be written as functions of $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$3 and $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$4 (Touzo et al., 2020).

A continuous-time formulation makes the filtration structure explicit. In the model of market makers facing customers who arrive at Poisson times and observe noisy signals of a finite-state Markov fundamental $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$5, the market maker does not observe the latent state or potential arrivals directly; he observes only actual buy and sell trades, represented by counting processes $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$6 and $$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$7. His filtration is

$$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$8

and the posterior state probabilities

$$P^{\text{ask}} = \mathbb{E}[\tilde V \mid \text{buy}], \qquad P^{\text{bid}} = \mathbb{E}[\tilde V \mid \text{sell}],$$9

solve a point-process filtering problem. At trade times, the ask and bid again equal conditional expectations of the latent value given the observed order-flow event and the current filtration (Kühn et al., 2012).

This continuous-time representation extends the order-flow component in two directions. First, buys and sells are informative through their arrival intensities: the probability of a buy at time $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$0 depends on the unobserved state. Second, the absence of trades is informative as well, because no-trade intervals also change posterior probabilities through the compensator terms in the filter (Kühn et al., 2012). A plausible implication is that the order-flow component should be interpreted more broadly as the entire filtration generated by order submissions and non-submissions, not only the sign of executed trades.

Point-process bridge constructions sharpen the same idea from another angle. In these constructions, the total order flow $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$1 is engineered so that, in its own filtration, it has the same law as the difference of two independent Poisson processes, while the terminal event $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$2 coincides with the insider’s information event. This yields an inconspicuous-trading equilibrium in which order flow looks like pure noise pathwise but still reveals the fundamental through its terminal distribution; as order size shrinks, the model converges weakly to the Kyle–Back equilibrium (Çetin et al., 2012, Li et al., 2013).

3. State-dependent allocation of the order-flow component across securities and trader types

Once multiple securities are introduced, the Glosten–Milgrom order-flow component ceases to be a single unconditional spread. In a multi-asset extension with an automated market maker that knows the joint distribution of asset values and conditions on the full vector of contemporaneous order flow, the price for security $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$3 becomes

$\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$4

where $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$5 is the multi-asset order-flow state. In the two-security case, the relevant states are $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$6, $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$7, $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$8, and $\tilde V_i \in \{p_i+r_i,\; p_i-r_i\},$9 (Gerig et al., 2010).

With positive value correlation, same-direction orders such as $\mathbb{E}[\tilde V_i]=p_i$0 and $\mathbb{E}[\tilde V_i]=p_i$1 are more likely to reflect informed trading than opposite-direction orders. The automated market maker therefore sets transaction prices farther from the unconditional mean $\mathbb{E}[\tilde V_i]=p_i$2 in same-direction states and closer to $\mathbb{E}[\tilde V_i]=p_i$3 in opposite-direction states. In this setting, the natural state-specific order-flow component is

$\mathbb{E}[\tilde V_i]=p_i$4

The original single-asset spread is thus replaced by a family of conditional adverse-selection premia indexed by the cross-section of order flow (Gerig et al., 2010).

A central result is that the unconditional spread remains

$\mathbb{E}[\tilde V_i]=p_i$5

but its incidence changes. Conditional on trader type,

$\mathbb{E}[\tilde V_i]=p_i$6

so informed traders face larger expected spreads and uninformed traders smaller ones after the automated market maker is introduced. The aggregate adverse-selection cost is unchanged for fixed $\mathbb{E}[\tilde V_i]=p_i$7, but it is reallocated toward orders and trader types that are more informative (Gerig et al., 2010).

The same model also shows that transaction prices become more efficient: the expected absolute pricing error $\mathbb{E}[\tilde V_i]=p_i$8 is smaller with the automated market maker than in the one-security benchmark. This follows because cross-asset order flow acts as an additional information source, allowing the liquidity provider to condition on a richer signal than own-asset order direction alone (Gerig et al., 2010).

The significance of this extension is conceptual as much as algebraic. In the classical model, all buys in a given asset look identical to the dealer. In the multi-asset model, order flows can be ranked by informativeness. The order-flow component therefore becomes a state-dependent object rather than a uniform spread.

4. Modern limit-order-book formulations: surprise, persistence, and transient impact

In electronic limit-order-book markets, the Glosten–Milgrom intuition survives, but the relevant state is no longer just a single buy-or-sell indicator. Given the full limit-order-book state and the full order-flow history, the price path is deterministic: $\mathbb{E}[\tilde V_i]=p_i$9 where $\gamma_i\in(0,1)$0 is the book state and $\gamma_i\in(0,1)$1 is the order-flow point process. Stochasticity enters because empirical models observe only a subset of $\gamma_i\in(0,1)$2, such as market orders or a single metaorder, and treat the remainder as noise (Lillo, 2021).

Within that setting, the Glosten–Milgrom order-flow component is often operationalized as the part of returns statistically attributable to signed order flow, especially its unexpected component. In the History Dependent Impact Model reviewed by Lillo, one has

$\gamma_i\in(0,1)$3

where $\gamma_i\in(0,1)$4 is trade sign, $\gamma_i\in(0,1)$5 is its linear predictor from past order flow, and only the surprise term $\gamma_i\in(0,1)$6 has permanent impact. In this interpretation, the permanent informational component of price changes is attached to innovations in order flow, while decaying propagator terms represent transient liquidity effects (Lillo, 2021).

Taranto, Bormetti, and Lillo provide an empirical counterpart. They document that market-order signs $\gamma_i\in(0,1)$7 exhibit slowly decaying autocorrelation, yet prices remain diffusive because impact is state dependent. Their reduced-form relation

$\gamma_i\in(0,1)$8

implies that predictable order flow has smaller average impact than surprising order flow. The key empirical mechanism is asymmetric liquidity: when a trade sign is more expected, the probability that the trade moves the price declines, largely because liquidity takers reduce aggressiveness and liquidity providers refill the opposite side of the book (Taranto et al., 2014).

This modern literature refines rather than overturns the classical concept. A raw order-sign series can be highly persistent without implying a proportionally large informational component of returns. What carries the Glosten–Milgrom-type content is typically the part of order flow that is not already anticipated by the market state or by recent order history (Taranto et al., 2014, Lillo, 2021).

A general operator framework for order-book dynamics makes the same separation explicit. The infinitesimal generator of the book can be decomposed into an order-flow generator and a deterministic clearing operator,

$\gamma_i\in(0,1)$9

so the stochastic order-flow component and the market-clearing component are analytically distinct. This suggests a direct modern analogue of the Glosten–Milgrom decomposition: one can isolate a pure order-flow mechanism and then study how the matching rule maps that mechanism into bid, ask, and mid-price dynamics (Cont et al., 2023).

5. Information-theoretic, market-design, and recent structural extensions

The order-flow component has also been reformulated in explicitly information-theoretic terms. In a thermodynamic analogy, the Glosten–Milgrom pricing rule is mapped to the optimal feedback rule of a Szilard engine: order flow becomes a sequence of noisy measurements of the binary fundamental, and the market maker’s posterior price corresponds to the optimal quasi-static position of the partition. In that formulation, the market temperature is

$$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$0

and the expected total gain of informed traders satisfies

$$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$1

A finite-horizon extension replaces $$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$2 by the mutual information in the observed order-flow sequence,

$$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$3

and derives a per-step inequality linking expected informed-trader gains to conditional mutual information carried by each order (Touzo et al., 2020, Carmier, 2022).

These results do not alter the market-microstructure meaning of the order-flow component; they recast it. The spread and informed-trader profits become bounded manifestations of entropy reduction through order flow. A plausible implication is that the order-flow component can be studied either as adverse-selection pricing or as the value of information extracted from a noisy measurement channel.

Recent work also shows how market design can attenuate the informational content of order flow itself. In a Glosten–Milgrom model with binary flip-noise in the market maker’s observation of trade direction, the observed signal is correct with probability $$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$4 and flipped with probability $$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$5. Under a committed Bayesian pricing rule, the spread becomes

$$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$6

so noisy direction observation linearly shrinks the order-flow component of the spread. The welfare decomposition yields a “privacy subsidy” of $$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$7 from the protocol’s liquidity pool to traders (Nakamura, 19 May 2026).

A distinct recent extension embeds a Glosten–Milgrom order-flow component into volatility forecasting for binary prediction markets. In that setting, per-event adverse-selection variance is approximated by $$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$8, where $$\mathbb{E}[\tilde V_i \mid B_i] = p_i+\gamma_i r_i,\qquad \mathbb{E}[\tilde V_i \mid S_i] = p_i-\gamma_i r_i,$$9 is the bid–ask spread, and trading volume proxies for the intensity of information-sensitive events. The resulting per-hour order-flow variance term is

$p_i+\gamma_i r_i$0

which is added to a Wright–Fisher deadline-resolution variance component. This uses spread as a measure of the size of order-flow-induced belief jumps and volume as a measure of their arrival rate (Xi et al., 9 Jul 2026).

Together these extensions show that the Glosten–Milgrom order-flow component is not confined to one interpretation. It can be a spread, a trader-specific adverse-selection burden, an information-theoretic work bound, a market-design object attenuated by privacy noise, or a structural variance component in binary markets. The common element is unchanged: order flow affects prices because it changes the posterior distribution of future payoffs.

6. Conceptual boundaries, common misconceptions, and synthesis

A first recurrent misconception is to identify the Glosten–Milgrom order-flow component with the entire bid–ask spread in all microstructure models. That identification is exact only in benchmark formulations that explicitly set aside inventory and order-processing costs. In the one-security benchmark discussed above, the spread is purely adverse selection for that reason, not by definitional necessity (Gerig et al., 2010).

A second misconception is to equate persistent order flow with permanent informational impact. Limit-order-book evidence indicates that highly predictable order flow can be associated with small marginal price effects once liquidity adapts. In empirical surprise-based impact equations, the effective Glosten–Milgrom component is attached to $p_i+\gamma_i r_i$1, not to raw $p_i+\gamma_i r_i$2 alone (Taranto et al., 2014, Lillo, 2021).

A third misconception is to treat the component as depending only on executed trade signs. Continuous-time filtering models show that the market maker’s information set can include both the arrival of buys and sells and the information content of no-trade intervals. Multi-asset models show that contemporaneous cross-asset order flow can refine the inference, and privacy mechanisms show that even the observation channel for trade direction can alter the component materially (Kühn et al., 2012, Gerig et al., 2010, Nakamura, 19 May 2026).

Across these variants, a stable definition emerges. The Glosten–Milgrom order-flow component is the part of prices, spreads, or price changes that is generated by conditioning on order flow as a signal about latent value. In the simplest binary model it is the spread $p_i+\gamma_i r_i$3. In richer environments it becomes a conditional object indexed by filtration, trader type, cross-asset state, surprise, or market design. What remains invariant is the underlying mechanism: liquidity providers require compensation, or prices adjust permanently, because observed order flow changes the posterior probability of underlying states.