2.10 Market Microstructure Game Theory

Basic Prerequisites: Probability & Measure Theory (Bayesian updating, conditional expectation) and Bayesian Statistics. (these are the folder-level prerequisites for pages 02–06; page 01 states its own, smaller, entry requirements)

1. Intuition & Practical Objective

A price is not a fact. It is the outcome of a game played every microsecond between three parties who want different things:

an informed trader (or several) who knows something about the fundamental value and wants to trade on it without revealing it;
noise/liquidity traders who trade for exogenous reasons and whose flow is the informed trader’s camouflage;
a market maker (or a competitive crowd of them) who must quote prices without knowing which counterparty it is facing, and therefore quotes conditionally - the ask is the expected value given that someone chose to buy from you, and the bid is the expected value given that someone chose to sell.

Every classical microstructure model is a specialization of that game. Kyle (1985) makes it a batch auction with a strategic insider and a linear price rule. Glosten–Milgrom (1985) makes it a sequential-trade Bayesian game with a competitive specialist and a two-point value. Brunnermeier–Pedersen (2005) adds a fourth player - a predator who trades against a forced liquidator. Schied–Zhang (2019) replaces the single insider with many strategic liquidators and asks what the Nash equilibrium of the execution game looks like.

The practical objective of this folder is exactly what a practitioner needs from that literature: when you trade, who is on the other side, what do they know, and what does the equilibrium price of that interaction look like? Adverse selection is not a parameter you sprinkle on top of a model - it is the equilibrium spread; impact is not a cost function you fit - it is Kyle’s λ; and your own optimal execution is not a private optimization problem - it is one player’s best response in a game.

This folder is the game-theoretic microstructure topic-folder for Pillar 2. As a hub, it gives you (a) the fast formula lookup below and (b) six sub-pages that walk from zero-knowledge intuition, through Kyle and the Back continuous-time limit, through Glosten–Milgrom as a Bayesian game, into predatory trading and execution games, then the failure modes and the modern extensions.

The one-sentence essence. “The spread is a Bayesian equilibrium price of adverse selection, impact is Kyle’s $λ = \frac{1}{2} Σ_{0} / σ_{u}^{2}$ , and your execution schedule is only optimal given the schedule everyone else is running - it is a Nash object, not a private one.”

Scope note (vs the sibling folders). This folder is the strategic/equilibrium view - who knows what, who moves when, and what the equilibrium of that interaction is. For the scheduling of a given order under a fixed impact model, see Optimal Execution & Almgren–Chriss. For the measurement of impact and depth, see Market Impact & Depth. For the sequential-Bayesian market-making view in depth, see Adverse Selection & the Glosten–Milgrom Model. The two meet at the same λ: this folder asks why the equilibrium has that λ, the others ask how to measure it and what to do with it.

Primary verified sources: Kyle (1985); Glosten & Milgrom (1985); Back (1992); Holden & Subrahmanyam (1992); Brunnermeier & Pedersen (2005); Carlin, Lobo & Viswanathan (2007); Schied & Zhang (2019); Schied, Strehle & Zhang (2018); Cordoni & Lillo (2022); Cardaliaguet & Lehalle (2018); Huberman & Stanzl (2004); Hasbrouck, Empirical Market Microstructure (Ch 5–7); Foucault, Pagano & Röell, Market Liquidity (Ch 3). Every number below was re-executed and reproduced in pure-stdlib Python (see §3).

2. Mathematical Ground Truth & Derivations

Notation. $v$ = fundamental value; $Σ_{0} = Var [v]$ = prior value variance; $u$ = noise order flow, $σ_{u}^{2} = Var [u]$ ; $x$ = informed demand; $y = x + u$ = total signed order flow; $P$ = price; $λ$ = price impact (“Kyle’s lambda”); $1/ λ$ = market depth. For the Glosten–Milgrom game: $V \in {V_{L}, V_{H}}$ , belief $θ_{t} = P (V = V_{H} ∣ F_{t})$ , $π$ = share of traders who are informed. For execution games: $X$ = block to liquidate, $T$ = horizon, $N$ = number of intervals of length $τ = T / N$ , $n_{k}$ = shares traded in interval $k$ , $x_{k}$ = shares still held, $η$ = temporary-impact coefficient, $γ$ = permanent-impact coefficient, $λ_{risk}$ = risk aversion.

2.1 The Glosten–Milgrom game (1985). Arrivals of buy/sell orders are governed by

P (B ∣ V_{H}) = \frac{1 + π}{2}, P (B ∣ V_{L}) = \frac{1 - π}{2}, P (S ∣ \cdot) = 1 - P (B ∣ \cdot) .

A competitive maker quotes zero expected profit conditional on the direction of the trade:

A_{t} = E [V ∣ buy at t], B_{t} = E [V ∣ sell at t],

which evaluates to

A_{t} = \frac{V _{H} ( 1 + π ) θ _{t} + V _{L} ( 1 - π ) ( 1 - θ _{t} )}{( 1 + π ) θ _{t} + ( 1 - π ) ( 1 - θ _{t} )}, B_{t} = \frac{V _{H} ( 1 - π ) θ _{t} + V _{L} ( 1 + π ) ( 1 - θ _{t} )}{( 1 - π ) θ _{t} + ( 1 + π ) ( 1 - θ _{t} )}

and the Bayesian update after a buy is $θ_{t}^{+} = \frac{( 1 + π ) θ _{t}}{( 1 + π ) θ _{t} + ( 1 - π ) ( 1 - θ _{t} )}$ . At $θ = \frac{1}{2}$ everything collapses to the textbook form

A - B = π (V_{H} - V_{L}) .

Key structural fact: the quoted price is a martingale - $E [P_{t + 1} ∣ F_{t}] = P_{t}$ - exactly because $A$ and $B$ are conditional expectations. Verified to $0.00 \times 1 0^{0}$ drift in §3.

2.2 The Kyle (1985) auction. Prior $v \sim N (p_{0}, Σ_{0})$ , noise $u \sim N (0, σ_{u}^{2})$ , one strategic insider, competitive linear pricing $P = p_{0} + λ y$ . The equilibrium is

x = β (v - p_{0}), β = σ_{u}^{2} / Σ_{0}, λ = \frac{1}{2} Σ_{0} / σ_{u}^{2}, 1/ λ = 2 σ_{u}^{2} / Σ_{0}

with $Var [v ∣ P] = Σ_{0} /2$ (exactly half the private information is impounded, and the fraction is invariant to $σ_{u}^{2}$ ) and $E [insider profit] = \frac{1}{2} σ_{u}^{2} Σ_{0}$ .

2.3 N-auction Kyle (the discrete game). Repeat the auction $N$ times with equal noise variance $σ_{u}^{2}$ per round and prior variance $Σ_{n - 1}$ . The equilibrium is

Σ_{n} = \frac{1}{2} Σ_{n - 1} = Σ_{0} 2^{- n}, β_{n} = σ_{u}^{2} / Σ_{n - 1}, λ_{n} = \frac{1}{2} Σ_{n - 1} / σ_{u}^{2}, E [profit_{n}] = \frac{1}{2} σ_{u}^{2} Σ_{n - 1}

so information is released geometrically (half per auction) and the insider’s total expected profit converges:

n \geq 1 \sum \frac{1}{2} σ_{u}^{2} Σ_{0} 2^{- (n - 1) /2} = \frac{2 + 2}{2} σ_{u}^{2} Σ_{0} .

$N = 12$ reproduces this limit to $1.6%$ (§3).

2.4 The Kyle–Back continuous-time limit. You cannot get continuous time by simply sending $N \to \infty$ in §2.3 with a fixed total noise budget: with per-period noise variance $σ_{u}^{2} τ$ the first-period impact is $λ_{1} = \frac{1}{2} Σ_{0} / (σ_{u}^{2} τ) \to \infty$ . The correct continuous-time equilibrium - Kyle (1985, §2) resolved by Back (1992) - has a linear information release and a constant impact coefficient:

Σ (t) = Σ_{0} (1 - \frac{t}{T}), λ = \frac{Σ _{0}}{σ _{u}^{2} T}

with $P_{t} = p_{0} + λ (X_{t} + U_{t})$ , $d U_{t} = σ_{u} d B_{t}$ , and market depth $1/ λ = σ_{u} T / Σ_{0}$ . Why the form is forced: the filtering identity $d Σ_{t} = - Var (d P_{t} ∣ F_{t^{-}}) = - λ^{2} σ_{u}^{2} d t$ combined with $Σ (t) = Σ_{0} (1 - t / T)$ gives $λ^{2} σ_{u}^{2} T = Σ_{0}$ identically - and $\int_{0}^{T} λ^{2} σ_{u}^{2} d t = Σ_{0}$ says all the information is released by $T$ . Two corollaries: half the information is impounded at $t = T /2$ (the continuous analogue of the auction’s $Σ_{0} /2$ ), and $λ \propto T^{- 1/2}$ (a longer horizon makes the market deeper, not shallower).

2.5 Predatory trading (Brunnermeier–Pedersen 2005). A distressed trader must liquidate $X$ on a schedule ${n_{k}}$ ; a predator best-responds to that public schedule. With price $P_{k} = P_{0} - κ \sum_{j \leq k} (n_{j} + m_{j})$ (sign convention: selling depresses) and linear temporary cost $η$ per unit for everyone, the predator’s cash flow is

Π = k \sum m_{k} (P_{k - 1} - η m_{k}), k \sum m_{k} = 0,

which reduces to a concave quadratic. Its maximiser is

m_{k} = \frac{κ ( A ˉ - A _{k} )}{2 η - κ}, A_{k} = j < k \sum n_{j}, \overset{ˉ}{A} = \frac{1}{N} \sum_{k} A_{k}

with the interior-solution condition $η > κ /2$ . Since $A_{k}$ is increasing, $m_{k}$ is positive early and negative late: the predator sells ahead of the victim and buys back afterwards, collecting the permanent-impact rent it helped create.

2.6 Execution as a game (Schied–Zhang 2019). When $J$ strategic liquidators trade the same asset, each minimises its own cost holding the others fixed. In the linear-quadratic case the symmetric Nash equilibrium is the single-agent Almgren–Chriss solution with an inflated temporary-impact coefficient: sharing one liquidity pool with a counterpart who trades $n_{k}^{j}$ makes agent $i$ ‘s marginal temporary cost $η (2 n_{k}^{i} + n_{k}^{j})$ , i.e.

η_{eff} = \frac{3}{2} η (two symmetric players) ⟹ κ_{eff} = λ_{risk} σ^{2} / η_{eff} = \frac{2}{3} κ

so competition for liquidity makes each player slower, not faster - the opposite of the naive intuition, and the reason “everyone liquidates at once” is expensive. Schied & Zhang (2019) and Cordoni & Lillo (2022) show the same structure in continuous time, where the equilibrium can become unstable (oscillating strategies, non-existent high-frequency limit) when the temporary-cost parameter is small relative to cross-impact.

2.7 Competing informed traders (Holden–Subrahmanyam 1992). Replace the single insider by $K$ insiders with aggregate informed intensity $K β$ . One round of Bayesian updating gives

Σ^{'} = \frac{Σ _{0} σ _{u}^{2}}{( K β ) ^{2} Σ _{0} + σ _{u}^{2}}, λ_{K} = \frac{K β Σ _{0}}{( K β ) ^{2} Σ _{0} + σ _{u}^{2}}

$K = 1$ recovers $Σ_{0} /2$ and $λ$ ; $K \to \infty$ drives $Σ^{'} \to 0$ and $λ_{K} \to 0$ . More insiders $\Rightarrow$ faster revelation $\Rightarrow$ a deeper market.

Quick-Reference Lookup - the fast facts of this folder (all reproduced by the stdlib engine in §3):

Quantity	Formula	Verified check
GM zero-profit spread ( $θ = \frac{1}{2}$ )	$A - B = π (V_{H} - V_{L})$	$π = 0.2, V_{H} = 101, V_{L} = 99 \Rightarrow 0.4000$
GM posterior after a buy	$θ^{+} = \frac{( 1 + π ) θ}{( 1 + π ) θ + ( 1 - π ) ( 1 - θ )}$	$θ = \frac{1}{2}, π = 0.2 \Rightarrow θ^{+} = 0.600000$
GM price is a martingale	$E [P_{t + 1} ∣ F_{t}] = P_{t}$	one-step mean $100.000000$ , drift $0.00 \times 1 0^{0}$
GM toxicity of a fill	$P (informed ∣ buy) = \frac{2 π θ}{( 1 + π ) θ + ( 1 - π ) ( 1 - θ )}$	$θ = \frac{1}{2} \Rightarrow 0.2000$
GM spread is state-dependent	$A (θ) - B (θ)$	peaks at $θ = \frac{1}{2}$ : $0.4000$ ; $θ = 0.95$ : $0.0785$
Kyle insider demand	$x = β (v - p_{0}), β = σ_{u}^{2} / Σ_{0}$	$Σ_{0} = 4, σ_{u}^{2} = 1 \Rightarrow β = 0.5000$
Kyle’s lambda	$λ = \frac{1}{2} Σ_{0} / σ_{u}^{2}$	$\Rightarrow λ = 1.0000$ , depth $1/ λ = 1.0000$
Kyle info impounded	$Var [v ∣ P] = Σ_{0} /2$	$2.0000$ (exactly half, independent of $σ_{u}^{2}$ )
Kyle insider profit	$E [π] = \frac{1}{2} σ_{u}^{2} Σ_{0}$	$1.0000$
Discrete $N$ -auction Kyle	$Σ_{n} = Σ_{n - 1} /2, β_{n} = σ_{u}^{2} / Σ_{n - 1}, λ_{n} = \frac{1}{2} Σ_{n - 1} / σ_{u}^{2}$	$λ_{1} = 1.0000, λ_{2} = 0.7071, λ_{3} = 0.5000$
Total insider profit ( $N \to \infty$ )	$\frac{2 + 2}{2} σ_{u}^{2} Σ_{0}$	$N = 12$ gives $3.3609 \to 3.4142$
Naive continuum is ill-posed	$λ_{1} = \frac{1}{2} Σ_{0} / (σ_{u}^{2} τ) \to \infty$	$N = 1 0^{4} \Rightarrow λ_{1} = 100.000$
Kyle–Back continuous time	$λ = Σ_{0} / (σ_{u}^{2} T), Σ (t) = Σ_{0} (1 - t / T)$	$T = 1 \Rightarrow λ = 2.0000$ ; $λ^{2} σ_{u}^{2} T = 4.0000 = Σ_{0}$
Half-information time (continuous)	$t^{⋆} = T /2$	$Σ (T /2) = 2.0000 = Σ_{0} /2$
Predator’s front-run schedule	$m_{k} = κ (\overset{ˉ}{A} - A_{k}) / (2 η - κ)$	$κ = 1, η = 0.6 \Rightarrow m_{1} = + 2.188, m_{8} = - 2.188$ , $\sum m_{k} = 0$
Cost of predation	victim loss / predator profit	loss $3.2812$ , predator profit $1.6406$ , deadweight $1.6406$
Victim’s best response	front-load (speed up)	victim loss $3.2812 \to 1.7175$ as it front-loads
Two-player execution Nash	$η_{eff} = \frac{3}{2} η$	quarters $46.81/26.18/15.70/11.31%$ = AC with $1.5 η$
$K$ competing insiders	$Σ^{'} = \frac{Σ _{0} σ _{u}^{2}}{( K β ) ^{2} Σ _{0} + σ _{u}^{2}}$	impounded: $K = 1 : 50%$ , $K = 2 : 80%$ , $K = 5 : 96.15%$
No-manipulation constraint	$γ + δ \geq 1$ (Gatheral/Huberman–Stanzl)	admissibility test on any impact/decay kernel

3. Computational Implementation - the game-theory core engine

Stdlib only (math, random), fully deterministic. It reproduces every verified number in the lookup table above: the GM spread, the GM martingale property, the Kyle auction closed forms, the N-auction recursion and its profit limit, the ill-posedness of the naive continuum, Back’s continuous-time information identity, the predator’s optimal front-run schedule, the two-player execution Nash, and the $K$ -insider information release.

The engine continues with the discrete Kyle recursion, the continuum ill-posedness, the Back information identity, the predator’s schedule, the two-player Nash and the $K$ -insider release:

One cross-check that matters: the shared-liquidity Nash equilibrium reproduces the single-agent Almgren–Chriss trajectory with $η_{eff} = 1.5 η$ to five decimal places in all four quarters, which is precisely the $η_{eff} = \frac{3}{2} η$ prediction of §2.6. The solver is therefore solving the game, not a mis-specified QP.

4. Failure Modes & First-Principles Breakdowns

Hub signposts - the full analysis lives in 05 - Failure Modes & Practice. One line each:

The equilibrium is only as good as the payoff it assumes. Glosten–Milgrom’s spread $π (V_{H} - V_{L})$ is an equilibrium of a game, not a fact about the world: mis-specify $π$ and the maker does not “earn a bit less”, it loses ~0.2 $/trade with certainty (verified in §3 of page 05).
The single-insider Kyle model is a fiction. Real markets have many informed participants; with $K$ insiders the price reveals information at rate $K β$ and λ collapses toward zero - calibrating a single-insider λ on a multi-insider market over-states impact.
The naive continuous-time limit does not exist. Discretising Kyle and taking $N \to \infty$ with a fixed noise budget sends λ to infinity; only Back’s (1992) linear-release equilibrium is well-defined. Any “high-frequency limit” claim must say which limit.
Strategic equilibria can be unstable. Schied & Zhang (2019) / Cordoni & Lillo (2022): below a threshold on the temporary-cost parameter the transient-impact execution game has wildly oscillating strategies and no well-behaved high-frequency limit - the equilibrium exists mathematically but is not a description of a market.
Predation is real but not identified by these models. The predator’s schedule depends on the victim’s schedule being public; in reality schedules are hidden and randomised, and the empirical literature struggles to separate predation from momentum and news.
Adverse selection is estimated, not observed. Practical proxies (PIN, VPIN) are noisy and contested (Easley–Kiefer–O’Hara–Paperman 1996; Easley–López de Prado–O’Hara 2012; Andersen & Bondarenko 2014), so the “equilibrium” π you plug in is itself a guess.

5. References

Kyle, Albert S. - “Continuous auctions and insider trading,” Econometrica 53(6), 1315–1335 (1985). The anchor of the whole folder: the strategic insider, the linear price rule, $λ$ , and the $Σ_{0} /2$ half-information result.
Glosten, Lawrence R.; Milgrom, Paul R. - “Bid, ask and transaction prices in a specialist market with heterogeneously informed traders,” Journal of Financial Economics 14(1), 71–100 (1985). Adverse selection as a sequential Bayesian game; zero-profit quotes are conditional expectations.
Back, Kerry - “Insider trading in continuous time,” Review of Financial Studies 5(3), 387–409 (1992). The correct continuous-time limit: linear information release, constant λ = √(Σ₀/(σ_u²T)), and the founding use of filtering in microstructure.
Holden, Craig W.; Subrahmanyam, Avanidhar - “Long-lived private information and imperfect competition,” Journal of Finance 47(1), 247–270 (1992). Many informed traders ⇒ faster revelation ⇒ deeper market; the antidote to single-insider calibration.
Brunnermeier, Markus K.; Pedersen, Lasse Heje - “Predatory trading,” Journal of Finance 60(4), 1825–1863 (2005). Predators sell alongside a distressed liquidator and buy back later; price overshoot, and illiquidity precisely when liquidity is needed.
Carlin, Bruce I.; Lobo, Miguel Sousa; Viswanathan, S. - “Episodic liquidity crises: Cooperative and predatory trading,” Journal of Finance 62(5), 2235–2274 (2007). The experimental/OECD-bond-market companion: when cooperation among potential predators breaks down.
Schied, Alexander; Zhang, Tao - “A market impact game under transient price impact,” Mathematics of Operations Research 44(1), 102–121 (2019). The canonical multi-agent execution game; source of the instability threshold.
Schied, Alexander; Strehle, Elias; Zhang, Tao - “High-frequency limit of Nash equilibria in a market impact game with transient price impact,” SIAM Journal on Financial Mathematics 8(1), 589–634 (2018). When the continuous-time limit of the execution game does (and does not) exist.
Cordoni, Francesco; Lillo, Fabrizio - “Instabilities in multi-asset and multi-agent market impact games,” Annals of Operations Research (2022). Extends Schied–Zhang to many agents and assets; the scaling of impact with the number of traders is what decides stability.
Cardaliaguet, Pierre; Lehalle, Charles-Albert - “Mean field game of controls and an application to trade crowding,” Mathematics and Financial Economics 12(3) (2018). *The mean-field limit of the execution game
Huberman, Gur; Stanzl, Werner - “Price manipulation and quasi-arbitrage,” Econometrica 72(4), 1247–1275 (2004). No-dynamic-arbitrage: the constraint that makes an impact/decay game admissible at all.
Easley, David; Kiefer, Nicholas M.; O’Hara, Maureen; Paperman, Joseph B. - “Liquidity, information, and infrequently traded stocks,” Journal of Finance 51(4), 1405–1436 (1996). *PIN
Easley, David; López de Prado, Marcos; O’Hara, Maureen - “Flow toxicity and liquidity in a high-frequency world,” Review of Financial Studies 25(5), 1457–1493 (2012). VPIN: the high-frequency toxicity proxy.
Andersen, Torben G.; Bondarenko, Oleg - “VPIN and the flash crash,” Journal of Financial Markets 17, 1–46 (2014). *The methodological takedown of VPIN
Biais, Bruno; Glosten, Lawrence; Spatt, Chester - “Market microstructure: A survey of microfoundations, empirical results, and policy implications,” Journal of Financial Markets 8(2), 217–264 (2005). The map of the whole field.
Hasbrouck, Joel - Empirical Market Microstructure (2007)
Foucault, Thierry; Pagano, Marco; Röell, Ailsa - Market Liquidity: Theory, Evidence, and Policy (2013)

6. Connected Graph Bridges

Foundational base: Probability & Measure Theory (conditional expectation, martingales) · Bayesian Statistics (the GM recursion) · Stochastic Calculus (Back’s filtering argument) · Calculus & Optimization (the quadratic execution games)
Sibling in-pillar: Optimal Execution & Almgren–Chriss (the non-strategic schedule that the games generalise) · Execution Algorithms: VWAP, TWAP, POV · Market Microstructure & Order Types · Queue Position & Fill Probability
Equilibrium-twin in Pillar 6: Market Impact & Depth (Kyle λ, the square-root law) · Adverse Selection & Glosten–Milgrom (the GM game in full) · Toxic Order Flow & VPIN (estimating π from flow) · Limit Order Book Mechanics
Market-making partner: Avellaneda–Stoikov & Optimal Quoting (the maker’s inventory game - the other half of the quote)
Risk & portfolio: Liquidity Risk & Funding (L-VaR) · Systemic Risk & Aggregation (predation as a spillover channel) · Constraints & Transaction Costs

Beginner: start at 01 · Practitioner: start at 05

Kwant Atlas

Explorer

2.10 Market Microstructure Game Theory

1. Intuition & Practical Objective

2. Mathematical Ground Truth & Derivations

3. Computational Implementation - the game-theory core engine

4. Failure Modes & First-Principles Breakdowns

5. References

6. Connected Graph Bridges

2.10.1 From Zero

2.10.2 Kyle (1985) and the Back Continuous-Time Limit

2.10.3 Glosten–Milgrom

2.10.4 Predatory Trading and Games Among Competing Liquidators

2.10.5 Failure Modes and Desk Practice

2.10.6 Advanced Extensions