Price impact tied to the depth within one tick of the mid gives a desk something it can measure. Aggregate market orders consume that depth, and the mid moves by κ times its relative change. Liu and Firoozi build their execution game around this mechanism. Their equilibrium proof, however, covers only a clipped market over short horizons.
The book comes before the price
Liu and Firoozi formulate a mean field game (MFG): each trader is individually negligible and responds to the average trading rate. They solve the limiting game through a fixed point, then show that it approximates a Nash equilibrium for finite N. The distinctive step is to pass aggregate flow through the order-book densities before it affects price. The paper contains theorems and one simulated example. It uses no real limit order book data and provides no calibration.
In most execution games, reduced-form impact lets the average trading rate move price directly. Here, bid and ask densities intervene. Each is a function of relative price and solves a stochastic partial differential equation (SPDE) extending the Cont and Müller limit order book model. A second-derivative term spreads orders through nearby cancellation and replacement; a first-derivative term draws them toward the mid. A negative zeroth-order term removes cancelled volume. Stochastic submission rates and infinite-dimensional multiplicative noise complete the book dynamics. Depth is the volume on each side within δ of the mid, where δ represents the tick. The mid-price obeys dF = κ(dD^b/D^b − dD^a/D^a).
Only net market-order imbalance connects traders to the book. The authors illustrate the scaling problem with 100 traders: 60 buy one unit and 40 sell one unit. That leaves an imbalance of 20 against a population average of 0.2. They scale initial book density with N and assume that the order-flow maps are homogeneous of degree one. With those assumptions, aggregate flow produces the same price, almost surely, as average flow in a per-trader system. Impact per unit of average flow thus remains unchanged as N grows. The authors describe their approach candidly, proceeding "heuristically to derive a system that is amenable to the MFG methodology while retaining a reasonable representation of the underlying market dynamics."
The trader's objective is familiar. Temporary impact is a·v; running and terminal inventory penalties are ϕq² and ξq². The optimal trading rate takes the linear-feedback form v* = −(F + Pq + φ)/(2a). P solves a Riccati equation, while φ solves a backward SDE driven by book noise. Equilibrium requires the assumed average rate to match the expected best response conditional on that noise. The authors prove an ε_N-Nash equilibrium for the N-player game, with ε_N = O(N^{-1/2}). The average trading rate and mid-price both converge to their limits at C/N in mean square.
What survives the proof?
The first requirement is regularity. To prove that depth is an Itô process, the authors need the density in an interpolation space with θ > 3/4. Their assumptions, 1/4 < α < 1/2 and a noise bound ρ < 1 − 2α, deliver that regularity. Remark 2 supplies an alternative when it fails: smooth the indicator 1_[0,δ] to obtain an approximate depth SDE.
The truncation has a more direct market consequence. Depth is clipped to [C_L, C_U] in the price equation, bounding impact per unit of net flow between κ/C_U and κ/C_L. The authors also clip the book-drift term Λ and project price volatility onto a ball of radius R. They acknowledge that this market "does not capture extreme market conditions, such as liquidity crises, extreme volatility, or very rapid changes in liquidity." C_L caps impact precisely when a thin book makes execution most expensive.
The horizon is restricted too.
Existence and uniqueness follow from contraction condition (4.86), which holds for small T. The authors explicitly limit their unique-solution result to all sufficiently small T. They connect that limit to truncation, introduced "for technical reasons", which "ensures the analytical solvability of the mean field game on a small time interval." Their sole numerical example nevertheless uses T = 5 and the untruncated market, beyond the theorem's guarantee. Positivity is another assumption: Remark 1 sketches conditions from an outside positivity result without imposing them on the densities.
Solver convergence, parameters without units
For the numerical example, the authors collapse the book into two scalar depth processes and leave price untruncated. They argue that sufficiently wide bounds bring the truncated and untruncated models arbitrarily close, with results "barely affected." The constants in their proof complicate that argument. The price-stability coefficient contains √δL/C_L, so widening the bounds by lowering C_L raises C_F(T). As C_F(T) rises, the horizon for which (4.86) guarantees a fixed point contracts. We did not find an evaluation of (4.86) at the example's T = 5.
The final Picard error was 5.64×10^-7. Average relative least-squares error in the backward regression was about 4.4×10^-3. These are solver diagnostics, and the authors' description of them as "numerical evidence" for a fixed point makes the appropriately limited claim.
The example starts at F̄_0 = 50, with depths of 40 reverting toward c = 50 at speed 1.5. Buying can depress the ask-side reversion level by as much as d = 48, bringing it down to 2. Each side has σ = 2; the remaining settings are κ = 7, q_0 = 5, a = 1 and penalties γ = α = 0.03. The horizon is T = 5, and units are unstated. We found no comparison of execution costs against TWAP or a reduced-form Almgren and Chriss benchmark.
Left to the implementer
- Choose δ, L and the covariance operators Q_a and Q_b. None is estimated.
- Specify the flow map G. Using the positive part of net flow, as the example does, discards gross volume by construction.
- Set C_L, C_U and R, which the theory requires and the example omits.
The notation also asks α to serve three purposes: a negative operator coefficient, an interpolation index between 1/4 and 1/2, and a terminal penalty of 0.03 in the example.
Implementation needs different data
We could not backtest this. We have daily and one-minute OHLCV bars; the mechanism requires bid and ask depth within one tick, together with signed market-order flow. A depth proxy made from bar ranges and volume would put a different impact model under the same name.
Testing this mechanism requires tick-level book snapshots over the full price window, signed trades, and cancel-and-replace events. Those observations are needed to estimate the diffusion, drift and decay coefficients, along with the replenishment speed fixed here at 1.5. The replenishment estimate deserves particular care. In the example, reversion speed alone determines how quickly depth depleted by aggregate buying returns toward c = 50.
The O(N^{-1/2}) approximation comes with unquantified constants. The paper therefore offers no guide to the N at which that approximation becomes useful. A calibration that found a stable κ across depth regimes, over a horizon where (4.86) actually holds, would move this work from interesting to usable.