Chapter 7Conclusion: The Information Wedge Across Environments

Every equilibrium in this dissertation is a set of kernels against the same primitive shocks. Under perfect information the kernels are the ordinary impulse responses. The chapters changed what players could see, and the coordinates never moved.

The games have two feedback channels: the physical one, and the one that runs through what other players learn.

The noise-state representation keeps the second channel by writing private information as conditional estimates of the primitive shocks while retaining endogenous signals, private histories, and individual impact. In the LQG case every object of the calculus is a deterministic kernel, and the information wedge, a term in the state-price equation, prices the effect of changing another player’s posterior.

The chapters vary what players see of the shocks. Monitoring varies what they know about a deviation’s origin. A movement a player observes may be filtered as noise or recognized as another player’s deviation. Chapter 6 prices both ways of responding, and the classical closed-loop game returns at the corner where every player sees the shocks and every player is privy.

The wedge and the accompanying equilibrium results come from the finite-horizon benchmark: Theorem 1.8 derives the belief-price representation, Corollary 1.11 gives closure over the full admissible class, and Proposition 1.12 solves unilateral best responses globally. Short-horizon equilibrium follows in Theorem A.7. The two-player benchmark separates statistical benefits from strategic costs and makes allocating signal precision a policy instrument.

The same object appears outside the baseline game. In identical-interest games the wedge prices the effect of shifting teammates’ posteriors and so their future controls (Section 1.4.1). Delayed public signals add to it a birth term and a jump in the belief price at the date the news arrives, and the stationary equations carry the same terms in lag coordinates. In Kyle–Back trading the wedge becomes a belief price on the market maker’s and the other traders’ filters, since order flow changes current prices and the future filtering environment (Theorems 4.6 and 4.10). On finite symmetric local markets, the graph symmetry reduces policy and filter kernels to equivalent-position coordinates (Proposition 5.7). Asymmetric monitoring splits the feedback into a naive channel, in which the naive player filters the deviation as noise (this is the wedge), and a privy channel, in which the privy player responds to the deviation directly, knowing who made it (Definition 6.7, Proposition 6.9).

7.1Directions beyond the dissertation

7.1.1Closure

A noise-state linear fixed point is already a Nash equilibrium against arbitrary admissible deviations, but whether every equilibrium takes this form is open. Proving it means allowing strategies with random martingale-representation coefficients, writing the induced nonlinear filters through the Kushner-Stratonovich equations, and showing that equilibrium forces those coefficients to collapse to deterministic kernels. The continuous-time martingale representation theorem gives a handle that is absent in the discrete setting where Witsenhausen’s counterexample [116] applies; a Lyapunov argument in the variance of the representation coefficients is the natural route.

7.1.2Computation and estimation

One route to computing the kernels adapts high-dimensional PDE methods, in the spirit of Han’s work [50], to the kernel HJB or the forward-backward kernel system. A different route approximates the noise-state itself. The filtering equations reweight old sources through kernels and project them into a current decision, the operation a linear attention layer performs. Linear transformers are a natural approximating class, with the noise-state as the target filter the trained network should recover. On the stationary system of Chapter 3, however, networks trained on the continuum residuals converged to spurious kernels while reporting small residuals (Figure 3.6).

On the dynamic-programming side, lifting the lag dynamics of Section A.2.3 to a fixed Hilbert space and taking a quadratic ansatz should recover the belief prices as off-diagonal Riccati blocks, and would likely give the CARA extension of the appendices a cleaner positivity condition than its operator inversions.

For empirical work the problem becomes estimation. Once the kernels are known the observed signal law has a Gaussian likelihood. But the kernels are endogenous equilibrium objects, so the estimator must either fit parameters while solving the kernel fixed point or learn a fast surrogate for that fixed point.

There is an easier road than fitting kernels. Across the chapters the strategic effects live in the means: the pooling gains of Chapter 1 are mostly mean-path, in Chapter 5 firms hold back their orders by several times the filtering gap while responses to aggregate shocks barely move, and the quote’s inventory loading of Chapter 6 is a level. The near-invariance of the responses to aggregate shocks is camouflage: an econometrician who fit those responses would match them and miss a level distortion, one that moves whenever policy changes who sees whose actions. The natural design is levels against observability changes—post-trade transparency [20, 45], platform visibility, disclosure mandates—rather than kernel comparisons.

7.1.3Sources

The most immediate extension is endogenous information acquisition. Signal technologies are treated as fixed throughout, and Appendix A.1 takes a first step by isolating the obstruction. One clean version uses a Brownian sheet of independent measurement channels, with each player selecting a portion of the sheet. Acquiring a source then changes the equilibrium policy kernels, and so changes what other players expect the holder to do with what it learns.

Compound Poisson signal sources are a larger step than a change of noise distribution. Throughout the dissertation each player knows the precision of their channels, so a source’s informativeness is fixed and the randomness lies in its realizations. A Poisson source with an uncertain arrival rate breaks this. Individual jumps may carry messages of known quality, but if the rate is unknown the value of the source as a whole is unknown. Signal quality itself becomes an inferred state, which is closer to how news, expert reports, and disclosures actually work. The filter is then generally infinite-dimensional, but that is not obviously the binding constraint; the question is whether the filter over primitive shocks stays stable and remains easier to approximate than the hierarchy it replaces.

7.1.4Networks

The most immediate extension of the graph chapter is the large-cycle limit of Conjecture 5.2. Replicated markets can also be linked endogenously, with a mean-field consistency condition on the distribution of market-level outcomes and possibly a few major markets of finite aggregate impact. Heterogeneous local environments would be aggregated over market types. In every case the aggregated object is a solved finite strategic environment, which distinguishes the construction from a graphon limit where local informational impact has already vanished.

7.1.5Markets

Chapter 6 makes the market maker strategic and computes the stationary market both ways, transparent and opaque. Existence remains open there, away from the corner cases and away from the $ρ \to 0$ degeneracy in which the market maker’s first-order condition cannot pin down how much the quote loads on inventory.

The delayed-public and monitored-deviation chapters suggest a separate extension in which attribution arrives after the action, with a trader’s order initially hidden in aggregate flow and a reporting rule naming it later. How players should respond on the day the order is named remains open.

In Rebonato’s parable [99], two identical volatility traders facing the same mispriced quote behave differently depending on whether the risk function marks their book to their own model or to the market consensus. In the Kyle–Back setting the consensus is the market maker’s price. A quote recognized as the market maker’s action is one the trader responds to directly, the way a privy player responds to an observed deviation in Chapter 6, while the trader’s private valuation remains a projection of primitive shocks through its noise-state. The gap between the trader’s valuation and the consensus mark is the performance measure in a principal–agent contract between the trader as agent and risk management as principal, with Sannikov-type continuation values encoding it [103]. The belief prices of Chapter 4 separately price order flow that moves the consensus. The contract, risk penalty, and strategic channel then jointly blunt how hard the trader trades on its private information.

7.1.6Noise traders with momentum

The strategic market maker above has a balance sheet it cares about, and Chapter 6 prices its quote deviations in two markets: a trader who sees the order flow recognizes a departure from the quote rule as the market maker’s, and a trader who sees only the quote filters it as noise. Either way the strategic market maker adds an inventory charge to the price. With Brownian noise that charge depends only on the inventory already held, since the noise adds variance but does not move the market maker’s forecast of future flow.

Noise traders with momentum do not need a strategic market maker. In Kyle–Back the informed traders already trade with long memory, since staying under the radar means trading slowly, and it is the noise flow that is Brownian. Retail herding has momentum. Fractional Brownian motion with $H > 1/2$ is the natural way to write this [81].

With both, the net order flow is a state the market maker manages, so a run of trend-chasing buying is a forecastable inventory problem rather than noise to be absorbed. The future of the noise now matters on its own, and everyone holds beliefs about it where Brownian noise gave them nothing to forecast. I expect the market maker’s conditional expectation of value and the price it posts to come apart, with the price reflecting inventory management against the flow the market maker expects.

A trader who can forecast the coming run, and the charge the market maker will add against it, buys ahead and sells into that charge. They buy to sell later. De Long et al. [34] have rational speculators front-running trend-chasing traders; here the trend-chasers’ persistence is a primitive. An order now moves the market maker’s forecast of flow as well as its estimate of value, so the trader manipulates through a second channel, and the wedge prices both. The round trip pays only if the market maker’s charge overshoots. The market maker forecasts the selling to come and prices it in, the traders buying ahead forecast the market maker, and how much of the rise survives would be a fixed point in the same kernels. The theorem to aim for is the market maker’s inventory cost against competition among the traders buying ahead.

As markets shift toward passive and flow-driven trading, the part of order flow that is not about fundamentals grows [41] and the market maker’s inventory problem becomes a larger part of the price. The model at the end has a strategic market maker who manages inventory, noise traders with momentum, and informed traders whose signals keep arriving.