Multi-Period Kyle-Type Model

• The multi-period Kyle-type model is a dynamic market microstructure framework where an insider trades alongside noise traders and a market maker using filtering to incorporate private information gradually.
• It employs advanced techniques such as fixed-point methods, Kalman filtering, optimal transport, and forward–backward systems to construct equilibrium and reveal true asset values.
• Extensions of the model address risk aversion, imperfect information, and multi-asset settings, highlighting its practical relevance and adaptability in modern financial markets.

A multi-period Kyle-type model is a dynamic market microstructure model in which an informed trader, noise or liquidity traders, and a market maker interact over several trading dates or on a continuous-time interval, while prices are set from aggregate order flow under rational expectations. In discrete time, trading occurs over periods $t=1,\dots,T$ or $n=1,\dots,N$; in continuous time, the same architecture is written on $[0,T]$ with aggregate order flow $Y_t=X_t+Z_t$ and prices of the form $P_t=\mathbb E[v\mid \mathcal F_t^Y]$ or closely related conditional expectations. Across the literature, the central objects are the insider’s dynamic strategy, the market maker’s filtering problem, and the endogenous path by which private information is incorporated into prices (Choi et al., 2015, Kühn et al., 2023, Back et al., 2020, Noh, 14 Jan 2026).

1. Canonical architecture and equilibrium notion

The canonical specification preserves the Kyle ingredients: an insider holds private information about a terminal or fundamental value, noise traders submit exogenous or partially endogenous order flow, and a market maker observes only total order flow and sets prices competitively. In discrete time, a representative formulation is

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$

where an insider and, in some models, a strategic rebalancer trade before competitive market makers update price beliefs. In the general discrete-time game, the market maker’s pricing system is

$$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$

on histories with positive probability (Choi et al., 2015, Kühn et al., 2023).

Continuous-time versions replace repeated auction rounds by a diffusion filtration. A baseline form is

$Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$

with the market maker observing $Y$ and setting

$P_t=\mathbb E[v\mid \mathcal F_t^Y].$

Under linear Markov pricing rules, many models write $n=1,\dots,N$0 or a state-augmented version such as $n=1,\dots,N$1 or $n=1,\dots,N$2. In the classical continuous-time benchmark, the insider gradually hides informed trading inside noise flow; in several extensions this is formalized by an “inconspicuous trading” condition such as

$n=1,\dots,N$3

This keeps the insider’s drift statistically hidden in the public filtration (Back et al., 2020, Chhaibi et al., 27 Jan 2025, Noh, 14 Jan 2026).

Equilibrium notions vary with the formal environment. The discrete target-trading model uses a linear Bayesian Nash equilibrium. The extensive-form finite-action formulation uses a sequential Kyle equilibrium, modeled on Kreps–Wilson sequential equilibrium, in which the insider is sequentially rational and the market maker’s beliefs are consistent with Bayes’ rule on-path and limiting beliefs off-path. Continuous-time formulations typically define equilibrium as a pair $n=1,\dots,N$4 or $n=1,\dots,N$5 satisfying insider optimality and rational pricing (Choi et al., 2015, Kühn et al., 2023, Bose et al., 2021).

2. Dynamic state variables, filtering, and public belief formation

The multi-period content of Kyle-type models is carried by filtering. The market maker does not observe the decomposition of total order flow and therefore updates beliefs from a hidden-state model whose dimension depends on the specification. In the target-execution model, the key latent state is

$n=1,\dots,N$6

the market makers’ belief about the rebalancer’s remaining demand. Prices and beliefs then evolve jointly through

$n=1,\dots,N$7

This creates autocorrelated order flow because

$n=1,\dots,N$8

The model is still linear-Gaussian, but the public state now includes beliefs about both value and future non-informational demand (Choi et al., 2015).

Continuous-time extensions enlarge the hidden state further. In the model with price-responsive traders, the state vector is

$n=1,\dots,N$9

where $[0,T]$0 and $[0,T]$1 are aggregate momentum-trader and contrarian-trader positions. Their dynamics are

$[0,T]$2

and order flow contains endogenous feedback through

$[0,T]$3

Because the system remains linear-Gaussian, the market maker’s conditional mean $[0,T]$4 follows a Kalman–Bucy filter,

$[0,T]$5

while the conditional covariance matrix satisfies a matrix Riccati equation (Noh, 14 Jan 2026).

Other continuous-time models use different state compressions. With imperfect information and insider risk aversion, the endogenous state is

$[0,T]$6

and the price becomes Markovian in $[0,T]$7: $[0,T]$8 In the stochastic-liquidity model, the market maker observes $[0,T]$9 and works with a transformed state $Y_t=X_t+Z_t$0, with

$Y_t=X_t+Z_t$1

so that conditional laws remain Gaussian after the appropriate innovation transformation (Chhaibi et al., 27 Jan 2025, Ekren et al., 2022).

A more general filtering perspective is developed in the Monge–Kantorovich approach. There the hidden state is $Y_t=X_t+Z_t$2, with conditional density

$Y_t=X_t+Z_t$3

and $Y_t=X_t+Z_t$4 solves the Kushner filtering SPDE. Reversing time yields a BSPDE with terminal condition induced by the optimal coupling of measures. In the path-dependent formulation, the price is a nonanticipative functional $Y_t=X_t+Z_t$5, and functional Itô calculus supplies the relevant horizontal and vertical derivatives $Y_t=X_t+Z_t$6 and $Y_t=X_t+Z_t$7 (Chhaibi et al., 2022, Corcuera et al., 2020).

3. Equilibrium construction techniques

A distinctive feature of multi-period Kyle-type theory is that equilibrium construction is rarely a single backward induction. Instead, the literature uses fixed points, filtering equations, optimal transport, Riccati systems, PDEs, BSPDEs, and bridge constructions.

One family of results rewrites the insider’s dynamic problem as a terminal optimal transport problem. In the generalized model with dynamic information, inconspicuousness implies the terminal order flow constraint $Y_t=X_t+Z_t$8, and the equilibrium terminal problem becomes

$Y_t=X_t+Z_t$9

with surplus $P_t=\mathbb E[v\mid \mathcal F_t^Y]$0. The pricing rule is then built from the Kantorovich potential: $P_t=\mathbb E[v\mid \mathcal F_t^Y]$1 and any continuous finite-variation strategy that enforces

$P_t=\mathbb E[v\mid \mathcal F_t^Y]$2

is optimal. The dynamic part is recovered through filtering and a BSPDE whose terminal condition is the disintegration of the optimal coupling (Chhaibi et al., 2022).

Optimal transport also underpins several continuous-time Gaussian and multidimensional constructions. In the Gaussian model with imperfect insider information and risk aversion, the terminal law of the state variable $P_t=\mathbb E[v\mid \mathcal F_t^Y]$3 is pinned down by a transport map $P_t=\mathbb E[v\mid \mathcal F_t^Y]$4, and solvability reduces to a fixed point for a variance parameter $P_t=\mathbb E[v\mid \mathcal F_t^Y]$5. In the multidimensional Kyle–Back model with a risk-averse informed trader, equilibrium is built from a convex potential $P_t=\mathbb E[v\mid \mathcal F_t^Y]$6 satisfying

$P_t=\mathbb E[v\mid \mathcal F_t^Y]$7

equivalently a Monge–Ampère-type equation, together with a Fokker–Planck equation and a coupled PDE system. In the risk-neutral and risk-averse dynamic Kyle model with general finite-variance fundamentals, equilibrium prices are obtained from heat-flow smoothing of a Brenier potential $P_t=\mathbb E[v\mid \mathcal F_t^Y]$8: $P_t=\mathbb E[v\mid \mathcal F_t^Y]$9 The insider’s value function is then

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$0

These constructions place Kyle equilibrium inside convex duality and Monge–Kantorovich duality rather than only stochastic control (Chhaibi et al., 27 Jan 2025, Bose et al., 2021, Back et al., 2020).

A second family of methods uses forward–backward systems. In the price-responsive-trader model, the insider is restricted to linear inconspicuous strategies

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$1

and equilibrium is a coupled forward–backward Riccati problem. Forward, the covariance matrix $$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$2 evolves by the Riccati equation; backward, Pontryagin’s Maximum Principle yields an adjoint matrix $$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$3 with

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$4

Under weak feedback, the equilibrium can be written as a fixed point

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$5

and the Banach-space implicit function theorem gives local existence, uniqueness, and smooth dependence on the feedback parameters (Noh, 14 Jan 2026).

Discrete-time formulations employ different fixed-point tools. The finite-support extensive-form game reduces to a static generalized social system in the sense of Debreu. After $$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$6-perturbing insider mixed strategies,

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$7

Kakutani’s fixed point theorem yields perturbed equilibria, and a limit argument produces a sequential Kyle equilibrium for every finite Kyle game. The continuous-state single-period game requires approximation by discrete games together with compactness for Young measures and a Komlós-type argument because “standard infinite-dimensional fixed point theorems are not applicable” (Kühn et al., 2023).

A third line of work interprets equilibrium order flow as a bridge. In the classical continuous-time Kyle model,

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$8

so total demand is a Brownian bridge to the insider’s signal. With transaction costs, the exact bridge is replaced by a Schrödinger-type bridge, and the constrained entropy minimization problem over path laws yields the equilibrium drift

$$y_n=\Delta\theta_n^I+\Delta\theta_n^R+\Delta w_n,\qquad p_n=E[v\mid \sigma(y_1,\dots,y_n)],$$9

The corresponding bridge laws converge to the classical Kyle equilibrium as trading costs vanish (Acciaio et al., 7 May 2026).

4. Principal model extensions

The multi-period Kyle framework has been extended along several economically distinct dimensions.

Risk aversion changes both terminal pricing and the insider’s intertemporal trade-off. In one Gaussian continuous-time model, the terminal pricing condition becomes

$$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$0

rather than $$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$1, because the insider internalizes terminal inventory risk. In the multidimensional Kyle–Back model with exponential utility, the equilibrium exists for sufficiently small $$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$2, and the insider’s strategy takes the bridge-like form

$$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$3

With risk-averse market makers, the dynamic Kyle model generates lower liquidity, inventory-based risk premia, short-term reversals, and excess volatility; in the normal case the order flow and price dynamics become explicitly mean reverting (Chhaibi et al., 27 Jan 2025, Bose et al., 2021, Back et al., 2020).

Imperfect private information modifies the hidden state known to the insider. In the Gaussian model with imperfect information, the insider sees only a noisy signal $$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$4 with

$$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$5

so the equilibrium cannot be reduced to the classical fully informed conditional-expectation logic. The solution combines a change of measure under which utility becomes a martingale, an optimal transport characterization of the terminal state, and a filtering system for the market maker (Chhaibi et al., 27 Jan 2025).

Multi-asset and non-Gaussian formulations generalize the scalar “Kyle lambda” into matrix-valued price-impact geometry. In the multidimensional risk-neutral transport model,

$$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$6

is symmetric positive semidefinite. In the multidimensional Kyle–Back model with strongly log-concave terminal law $$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$7, equilibrium pricing is path dependent and nonlinear in general. This shifts the theory from scalar Gaussian filtering to matrix-valued and nonlinear transport/filtering systems (Back et al., 2020, Bose et al., 2021).

Strategic heterogeneity introduces agents with motives other than informed speculation. The dynamic rebalancing model adds a strategic portfolio rebalancer who must satisfy

$$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$8

and trades according to

$$S_t:(E_Y)^t\to [v_N,v_1],\qquad S_t(y_1,\dots,y_t)=\sum_{i=1}^N p_Y^\xi(i\mid y_1,\dots,y_t)\,v_i$$9

while the insider continues to trade on long-lived information. In the price-responsive-trader model, endogenous “noise” flow is decomposed into momentum and contrarian components, which alter both the observation equation and the speed of price discovery (Choi et al., 2015, Noh, 14 Jan 2026).

Liquidity-side extensions alter the law of noise trading or the market maker’s objective. In the stochastic-liquidity model,

$Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$0

so noise-trading volatility is stochastic and may be correlated with the volatility driver observed by the market maker. In the adaptive-agent discrete model, price impact evolves recursively via

$Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$1

and slow belief revision about $Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$2 creates endogenous volatility dynamics. In the market-maker-revenue model, the market maker minimizes

$Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$3

so bid-ask spread revenue enters directly into the pricing problem (Ekren et al., 2022, Vodret et al., 2022, Lehalle et al., 2021).

Horizon modifications alter the geometry of information revelation. In the random-horizon model, the announcement time $Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$4 is exponentially distributed and the price satisfies

$Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$5

In the stationary setup there is no terminal revelation of a single fundamental value; instead the asset pays a dividend stream and equilibrium price takes the propagator form

$Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$6

This replaces fixed-horizon learning by stationary filtering of persistent order flow and dividend signals (Çetin, 2016, Vodret et al., 2020).

5. Price discovery, liquidity, and dynamic implications

The central comparative-static object in most multi-period Kyle-type models is the path of posterior uncertainty about value. In discrete time this is often $Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$7; in continuous time it may be $Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$8. The price-responsive-trader model writes, for example,

$Y_t=X_t+Z_t,\qquad dZ_t=\sigma\,dW_t,$9

which directly links the speed of price discovery to insider intensity and to momentum/contrarian feedback. Under weak feedback, momentum feedback satisfies

$Y$0

so prices become more informative faster, and insider informational rents fall (Noh, 14 Jan 2026).

Dynamic liquidity patterns vary sharply across specifications. In the target-rebalancing model, equilibrium price impact $Y$1 is $Y$2-shaped rather than monotone, the insider’s expected trading intensity can be slightly $Y$3-shaped, the rebalancer’s trading is $Y$4-shaped, order flow becomes autocorrelated, and insider and rebalancer orders tend to become negatively correlated later in the day. The paper interprets this as a “symbiotic liquidity provision” effect (Choi et al., 2015).

Different insider objectives generate different information-release paths. In the discrete model with rational expected prices, a risk-averse insider front-loads trading and, as trading frequency approaches infinity, the model “acquires a strong-form efficiency.” A risk-neutral insider yields a discrete equilibrium that converges to Kyle’s continuous-time equilibrium: $Y$5 A risk-seeking insider delays revelation, with

$Y$6

This suggests that the temporal profile of adverse selection is not pinned down by rational pricing alone; it also depends on the insider’s preference ordering over guaranteed and risky profits (Gong et al., 2010).

Several extensions imply nonclassical dynamics for Kyle’s lambda and market depth. In the random-horizon model, lambda is a uniformly integrable supermartingale and, in the Bernoulli case, a potential with $Y$7. By contrast, in the stochastic-liquidity model both

$Y$8

are submartingales. The paper stresses that this is “surprisingly” true for market depth as well (Çetin, 2016, Ekren et al., 2022).

Some extensions connect Kyle dynamics to broader empirical phenomena. With risk-averse market makers, the dynamic optimal-transport model generates short-term reversals and excess volatility because prices include an inventory-risk component beyond the information content of order flow. With adaptive agents and time-varying noise-trade volatility, the sticky-expectation limit yields

$Y$9

a Kesten process that the paper presents as a microfoundation for GARCH-like volatility clustering. In the lognormal-fundamental stochastic-liquidity model, informed trading forces the conditional log-return to maturity to be Gaussian even though the price process itself has stochastic volatility (Back et al., 2020, Vodret et al., 2022, Ekren et al., 2022).

6. Existence, uniqueness, and scope conditions

Existence and uniqueness are not automatic in multi-period Kyle-type models. In the price-responsive-trader extension, if the feedback vector

$P_t=\mathbb E[v\mid \mathcal F_t^Y].$0

is sufficiently small, there exists a unique equilibrium $P_t=\mathbb E[v\mid \mathcal F_t^Y].$1 that is a smooth perturbation of the classical Kyle intensity. For stronger feedback, the paper identifies three failure modes: finite-time breakdown of the Riccati flow, loss of contraction of the equilibrium map, and instability of the Kalman filter. In that regime, multiplicity of equilibria and amplification effects can arise (Noh, 14 Jan 2026).

Discrete-time finite-support existence is broader but structurally different. The extensive-form model proves that every finite Kyle game has a sequential Kyle equilibrium, yet equilibria “exist in general only in mixed strategies and not in pure strategies.” In the single-period case, the insider’s order is monotone in the true value, but the equilibrium price function need not be monotone. This directly qualifies the common continuous-normal intuition that monotone informed demand necessarily induces monotone pricing (Kühn et al., 2023).

Some apparently general extensions turn out to be more restrictive in equilibrium than their formulation suggests. The path-dependent model allows prices to depend on the whole past path $P_t=\mathbb E[v\mid \mathcal F_t^Y].$2, but its main equilibrium theorems require terminal revelation, martingale aggregate order flow, and a collapse of marginal price impact to a spot-Markovian form: $P_t=\mathbb E[v\mid \mathcal F_t^Y].$3 The paper states that “genuinely path-dependent pricing rules are difficult to sustain in equilibrium” under its conditions. This suggests that allowing path dependence at the level of admissible functionals does not by itself guarantee a path-dependent equilibrium law (Corcuera et al., 2020).

The category “multi-period Kyle-type model” also requires care in classification. Some papers explicitly develop finite-horizon repeated trading or continuous-time dynamic learning, whereas others only reference multi-period Kyle in motivation. The market-maker-revenue paper is “primarily one-period,” and its “dynamics” are algorithmic iterations in the equilibrium solver rather than economic multi-period trading dynamics. By contrast, the random-horizon, stationary, adaptive-agent, and transport/filtering models all replace the one-shot setting with genuinely time-indexed learning, strategic interaction, or state evolution (Lehalle et al., 2021, Çetin, 2016, Vodret et al., 2020, Chhaibi et al., 2022).

Taken together, these results delimit the modern meaning of a multi-period Kyle-type model: a dynamic adverse-selection equilibrium in which private information is incorporated gradually through order flow, public beliefs evolve by filtering or its nonlinear generalizations, and equilibrium existence depends on the compatibility of insider incentives, market-maker inference, and the chosen state-space or bridge structure.