betaFeedback

WorksEconomics, Econometrics and Finance

Learning AI Agents, Market Makers, and Stress in Automated Markets

Abstract

Automated markets increasingly route order flow from learning trading agents through automated market makers (AMMs). We model trader archetypes as AI agents defined by their reward functions and analyze the feedback they create on a constant-product AMM. Herd-dominated and arbitrage-dominated markets of equal size have identical return distributions and turnover, so no rule based on volatility, value-at-risk, stress frequency, or volume can tell them apart; yet their stress episodes differ in length by a factor unbounded in the feedback-to-noise ratio β/sβ/s. Serial statistics identify the feedback sign from Θ((s/β)2log⁡(1/δ))Θ((s/β)^2\log(1/δ)) observations, and this rate is tight. When agents trade together, profit-seeking agents ride herds rather than lean against them, leading to all-or-nothing participation. Spread-earning market makers restore stability but make liquidity fragile, with discontinuous collapse and hysteresis; maker–taker pricing, not a uniform fee, is the effective instrument.

Introduction

A growing share of trading is done by automated, increasingly autonomous agents, and a growing share of matching, above all in digital assets, happens against automated market makers: deterministic pricing formulas such as the constant-product rule [1, 3]. Arbitrageurs, retail trend-followers, and front-runners [7] all trade in the same pools, and each leaves a different footprint on the frequency and persistence of stress. As AI agents take on these roles, proportionate regulation requires knowing which objectives stabilize a market and whether a regulator can determine this from data.

Two obstacles stand in the way. Behavior types are not labeled in public data, and the statistics regulators usually monitor (volatility, value-at-risk, the frequency of large losses, turnover) are properties of the one-period return distribution. We show that such statistics are provably blind to the property that matters most, identify what is not blind to it at an optimal rate, and analyze what learning agents and market makers do when they trade together. We take an objective-first view, following the empirical framework of [2]: each agent is defined by the reward it maximizes, so assumptions are explicit while strategy is emergent.

Our results

We model the pool’s log return as rt=It+ηtr_t = I_t + \eta_t, where ItI_t is the price impact of the period’s net order flow and ηt\eta_t is exogenous news with scale ss. Herd-followers trade in the direction of the last move, which we show is optimal under the herd-follower’s reward (Lemma 3.2). In the simplest case rt=βsgn⁡(rt−1)+ηtr_t = \beta\operatorname{sgn}(r_{t-1}) + \eta_t, where β>0\beta> 0 is net positive feedback (herding) and β<0\beta< 0 net negative feedback (arbitrage).

Volatility cannot see feedback (§ 4). Markets with feedback +β+\beta and −β-\beta have the same one-period return law and the same turnover, but their expected stress-episode lengths differ by G(c+βs)/G(c−βs)G\left(\frac{c+\beta}{s}\right)/G\left(\frac{c-\beta}{s}\right), which is unbounded in β/s\beta/s (Theorem 4.2). Therefore, every functional of the return marginal and turnover treats a destabilizing and a stabilizing market identically (Corollary 4.3). For Gaussian news with β=s\beta= s, episodes last 6.3 periods under herding and 1.2 under arbitrage (Figure 1a).

Figure 1. (a) Theorem 4.2: expected stress-episode length (c=0c = 0, Gaussian news) in a herd-dominated and an arbitrage-dominated market. The two have identical volatility and stress frequency at every ∣β∣|\beta|. (b) Theorem 6.1(v): expected down-run length against rider capital Xˉ\bar{X} (λ=s=1\lambda= s = 1, h=0.3h = 0.3, ϕ=0.2\phi= 0.2, σˉ=1\bar{\sigma} = 1). Market makers exit at Xˉc=0.51\bar{X}_{c} = 0.51 and return only below Xˉlo=0.21\bar{X}_{lo} = 0.21; in the shaded band both outcomes are stable.

Serial statistics see it at an optimal rate. Autocorrelation and the variance ratio have the sign of β\beta (Proposition 4.4). A sign-agreement test identifies sgn⁡(β)\operatorname{sgn}(\beta) with error δ\delta from O((s/β)2log⁡(1/δ))O((s/\beta)^2 \log(1/\delta)) observations, and no test using the full return path needs fewer than Ω((s/β)2log⁡(1/δ))\Omega((s/\beta)^2 \log(1/\delta)) (Theorem 4.5). A herd-follower’s own stress episodes, by contrast, last at most two periods on every path (Proposition 4.6), so strategy-level measurements see herders as brief sources of stress and miss their effect on market persistence.

Agents playing together (§ 5). On a path-independent pool, a trader’s own impact cancels from its round-trip payoff (Lemma 5.1); only trading ahead of predictable herd flow pays. Leaning against the herd is therefore strictly dominated, riding with it is a strategic complement, and under any sign-preserving learning dynamics, including replicator dynamics and softmax policy gradient, participation tips to zero or to full depending on the starting point; full participation strictly lengthens stress, and no uniform fee can induce stabilizing behavior (Theorem 5.2, Corollary 5.3).

Market makers (§ 6). Spread-earning market makers enter the price process exactly as negative riders, so persistence depends only on the net index z=w−σhz = w - \sigma h of rider capital minus absorbed flow. They supply the negative feedback fee-paying traders will not, but liquidity becomes fragile: when rider capital crosses a threshold Xˉc\bar{X}_c market makers exit at once and persistence jumps, and they return only below Xˉc−σˉhˉ\bar{X}_c-\bar{\sigma}\bar{h}, so the hysteresis band is exactly as wide as their absorption capacity (Theorem 6.1, Figure 1b). A maker rebate widens the stable region; a taker fee removes riders (Corollary 6.2).

Techniques. Three observations drive the analysis. (i) With symmetric noise, the return sign is a Markov chain whose flips are i.i.d. and independent of the current sign (Lemma 4.1), which decouples the marginal (a function of ∣β∣|\beta|) from the serial structure (a function of sgn⁡(β)\operatorname{sgn}(\beta)). (ii) On a path-independent pool, the own-impact part of a round trip telescopes away in expectation (Lemma 5.1), so payoffs depend on aggregate behavior only through one persistence probability qq. (iii) qq is monotone in each population’s net flow, which makes the interaction game one of strategic complements within groups and substitutes across groups [16], so equilibria, stability, and learning basins follow from monotone comparative statics.

Empirical connection. Section 7 relates the results to the companion study [2], which trains these archetypes with deep reinforcement learning on hourly cryptocurrency data. The theory explains its main strategy-level patterns and shows which of them do not survive interaction.

Related work

De Long, Shleifer, Summers, and Waldmann [8] show that rational speculators may amplify positive-feedback traders; Theorem 5.2 derives this for learning agents on an AMM, with equilibrium tipping and learning basins. Agent-based finance shows that heterogeneous traders produce fat tails and crashes [14]. Classical microstructure explains how liquidity providers set prices [13, 9] and absorb temporary imbalances [10]; liquidity can evaporate in spirals [4], as in the 2010 Flash Crash [12]. Our market-maker result gives a closed-form collapse threshold and hysteresis band. Constant-function market makers are analyzed in [3], arbitrage costs against AMMs in [17], and front-running on decentralized exchanges in [7]; our execution model follows frequent batch auctions [5]. Learning algorithms can reach unintended outcomes such as tacit collusion [6]; our learning results use sign-preserving population dynamics [18] and games with strategic complementarities [16]. The variance-ratio test of [15] is a serial statistic of the kind Theorem 4.5 shows a regulator needs.

Organization. Section 2 recalls the archetypes’ rewards and Section 3 gives the AMM feedback model. Section 4 proves the impossibility and detection results, and Sections 5 and 6 analyze agents trading together and market makers. Section 7 connects the results to data, Section 8 discusses implications for regulators, and Appendix A verifies every closed form by simulation.

Trader archetypes as reward functions

We recall the archetypes of [2]. Trading is a discrete exposure-control problem. At each step tt, the agent observes a feature vector oto_t and chooses an action in {sell,hold,buy}\{\text{sell},\text{hold},\text{buy}\}, which moves its exposure et∈{0,12,1}e_t\in\{0,\frac{1}{2},1\} by −Δ-\Delta, 00, or +Δ+\Delta with Δ=12\Delta=\frac{1}{2}. Exposure 00 is flat, 12\frac{1}{2} half invested, and 11 fully long; there is no short selling. The base reward is a volatility-normalized net return with a stress penalty:

rtbase=etRt+1−ctx∣et−et−1∣σ^t+ϵ−λstress1[stresst]et(1)r_t^{\mathrm{base}}=\frac{e_tR_{t+1}-c_{\mathrm{tx}}\lvert e_t-e_{t-1}\rvert}{\hat{\sigma}_t+\epsilon}-\lambda_{\mathrm{stress}}\mathbf{1}[\mathrm{stress}_t]e_t \tag*{(1)}

where Rt+1R_{t+1} is the next-period asset return, ctxc_{\mathrm{tx}} the proportional transaction cost, σ^t\hat{\sigma}_t trailing volatility, ϵ>0\epsilon> 0 a small constant, and stresst\mathrm{stress}_t a market crash-regime indicator. Each archetype modifies (1):

  • Rational arbitrageur. Maximizes risk-adjusted return with additional penalties on realized volatility and turnover. It acts conservatively and reduces exposure as risk rises.

  • Manipulator. A stylized proxy for front-running and wash trading. It receives a bonus κturn∣et−et−1∣\kappa_{\mathrm{turn}}|e_t-e_{t-1}| and a short discount horizon, which favors rapid position cycling.

  • Retail herd-follower. Mimics trend-chasing and fear of missing out. It receives a bonus for aligning exposure with recent momentum,

κmomsgn⁡(momt)(et−12).(2)\kappa_{\mathrm{mom}} \operatorname{sgn}(\mathrm{mom}_t)\left(e_t-\frac{1}{2}\right). \tag*{(2)}

Sections 3–6 ask what these objectives do to a market when the agents’ trades move prices; the companion study [2] trains them on market data.

A feedback model of an automated market

The AMM. A constant-product pool with reserves (x,y)(x,y) quotes price p=y/xp=y/x. Depositing dd units of yy moves the reserves to (x′,y+d)(x',y+d) with x′(y+d)=xyx'(y+d)=xy, so

log⁡p′−log⁡p=2log⁡(1+dy)=2dy+O(d2y2).\log p' - \log p = 2\log\left(1+\frac{d}{y}\right)=\frac{2d}{y}+O\left(\frac{d^2}{y^2}\right).

Because the pool is path-independent, a period’s orders aggregate into one net flow NtN_t, measured in units of the yy-reserve. We use the first-order impact λNt\lambda N_t with λ=2\lambda=2; Remark 4.7 treats the exact impact.

Order flow. Time is discrete, t=1,2,…t=1,2,\ldots, and rt=log⁡pt−log⁡pt−1r_t=\log p_t-\log p_{t-1} is the pool’s return. Each period brings exogenous news ηt\eta_t (for example, moves in an external reference market transmitted to the pool) and endogenous flow from two populations:

  • mass h≥0h\geq0 of herd-followers, each buying one unit if the last return was positive and selling one unit otherwise; and

  • mass α≥0\alpha\geq0 of negative-feedback traders, each trading one unit against the last return.

Writing β=λ(h−α)\beta=\lambda(h-\alpha) and Dt=sgn⁡(rt)D_t=\operatorname{sgn}(r_t), the pool’s return is

rt=βDt−1+ηt.(3)r_t=\beta D_{t-1}+\eta_t. \tag*{(3)}

Assumption 1. The ηt\eta_t are i.i.d. and independent of r0r_0, with ηt=sξt\eta_t=s\xi_t for s>0s>0, where ξt\xi_t has a continuous, strictly increasing distribution function GG that is symmetric about zero (G(−x)=1−G(x)G(-x)=1-G(x)). The initial sign D0D_0 is uniform on {±1}\{\pm1\}.

Definition 3.1 (Stress). For a threshold c≥0c\geq0, period tt is stressed if St=1[rt<−c]S_t=1[r_t<-c] equals one. A stress episode is a maximal run of consecutive stressed periods, and LL denotes its length. The stress frequency is F=P(St=1)F=\mathbb{P}(S_t=1), the volatility is Var⁡(rt)\sqrt{\operatorname{Var}(r_t)}, and the gross turnover is h+αh+\alpha units per period.

From rewards to feedback. The next lemma shows that the herd-follower’s reward induces the herd rule discussed above.

Lemma 3.2 (The herd reward induces sign-momentum). Let μt=E[Rt+1∣Ft]\mu_t = \mathbb{E}[R_{t+1} \mid\mathcal{F}_t], and suppose the agent is myopic (discount factor 0) with et−1∈(0,1)e_{t-1} \in(0,1). If

κmom>∣μt∣+ctxσ^t+ϵ+λstress,\kappa_{\mathrm{mom}} > \frac{|\mu_t| + c_{\mathrm{tx}}}{\hat{\sigma}_t + \epsilon} + \lambda_{\mathrm{stress}},

then the unique expected-reward-maximizing action under (1)–(2) is buy when mom⁡t>0\operatorname{mom}_t > 0 and sell when mom⁡t<0\operatorname{mom}_t < 0.

Proof. Write the action as j∈{−1,0,1}j \in\{-1,0,1\}, so e=et−1+j/2e = e_{t-1} + j/2, and let χt=sgn⁡(mom⁡t)\chi_t = \operatorname{sgn}(\operatorname{mom}_t). The expected one-step reward is eK−ctx∣j∣2(σ^t+ϵ)eK - \frac{c_{\mathrm{tx}}|j|}{2(\hat{\sigma}_t+\epsilon)} plus terms that do not depend on jj, where K=μtσ^t+ϵ−λstress1[stress⁡t]+κmomχtK = \frac{\mu_t}{\hat{\sigma}_t+\epsilon} - \lambda_{\mathrm{stress}}\mathbf{1}[\operatorname{stress}_t] + \kappa_{\mathrm{mom}}\chi_t. Relative to holding, buying gains 12(K−ctxσ^t+ϵ)\frac{1}{2}\left(K - \frac{c_{\mathrm{tx}}}{\hat{\sigma}_t+\epsilon}\right) and selling gains 12(−K−ctxσ^t+ϵ)\frac{1}{2}\left(-K - \frac{c_{\mathrm{tx}}}{\hat{\sigma}_t+\epsilon}\right). If χt=1\chi_t = 1, the hypothesis gives K≥κmom−∣μt∣σ^t+ϵ−λstress>ctxσ^t+ϵ>0K \ge\kappa_{\mathrm{mom}} - \frac{|\mu_t|}{\hat{\sigma}_t+\epsilon} - \lambda_{\mathrm{stress}} > \frac{c_{\mathrm{tx}}}{\hat{\sigma}_t+\epsilon} > 0. Therefore, buying beats holding, and buying also beats selling by K>0K > 0. If χt=−1\chi_t = -1, then −K≥κmom−∣μt∣σ^t+ϵ>ctxσ^t+ϵ-K \ge\kappa_{\mathrm{mom}} - \frac{|\mu_t|}{\hat{\sigma}_t+\epsilon} > \frac{c_{\mathrm{tx}}}{\hat{\sigma}_t+\epsilon}, so selling is the unique maximizer.

With a one-period momentum signal, a population of such agents emits a net flow hDt−1hD_{t-1}, the herd term in (3). In § 4, the negative-feedback mass α\alpha is a modeling device; § 5 asks whether any profit-seeking agent would choose to lean, and § 6 shows who does.

Volatility cannot see feedback

Lemma 4.1 (The sign process). Let q=G(β/s)q = G(\beta/s). Under Assumption 1, the flips At=DtDt−1A_t = D_tD_{t-1}, t≥1t \ge1, are i.i.d. with P(At=1)=q\mathbb{P}(A_t = 1) = q and independent of D0D_0. Consequently, (Dt)(D_t) is a stationary symmetric Markov chain with uniform marginals.

Proof. At=sgn⁡(Dt−1rt)=sgn⁡(β+Dt−1ηt)A_t = \operatorname{sgn}(D_{t-1}r_t) = \operatorname{sgn}(\beta+ D_{t-1}\eta_t). Conditional on Ft−1\mathcal{F}_{t-1}, the variable Dt−1ηtD_{t-1}\eta_t has the law of ηt\eta_t, by independence and symmetry. So AtA_t is independent of Ft−1\mathcal{F}_{t-1} with P(At=1)=P(ηt>−β)=G(β/s)\mathbb{P}(A_t = 1) = \mathbb{P}(\eta_t > -\beta) = G(\beta/s). Since Dt=D0∏k≤tAkD_t = D_0 \prod_{k \le t} A_k and D0D_0 is uniform, each DtD_t is uniform.

Theorem 4.2 (Volatility equivalence, persistence separation). Under Assumption 1, for every t≥1t \ge1 and every c≥0c \ge0:

(i) The law of rtr_t is 12Gs(⋅−β)+12Gs(⋅+β)\frac{1}{2}G_s(\cdot-\beta) + \frac{1}{2}G_s(\cdot+\beta), where Gs(x)=G(x/s)G_s(x) = G(x/s). It depends on β\beta only through ∣β∣|\beta|. In particular, Var⁡(rt)=β2+s2Var⁡(ξ1)\operatorname{Var}(r_t) = \beta^2 + s^2\operatorname{Var}(\xi_1) when ξ1\xi_1 has finite variance, and

F(β)=12[G(−c−βs)+G(β−cs)]=F(−β).F(\beta) = \frac{1}{2}\left[G\left(\frac{-c-\beta}{s}\right) + G\left(\frac{\beta-c}{s}\right)\right] = F(-\beta).

(ii) Stress-episode lengths are i.i.d. geometric with P(L>k)=G(β−cs)k\mathbb{P}(L > k) = G\left(\frac{\beta-c}{s}\right)^k, so

E[L](β)=1G(c−βs),\mathbb{E}[L](\beta) = \frac{1}{G\left(\frac{c-\beta}{s}\right)},

which is strictly increasing in β\beta.

(iii) For β>0\beta> 0,

E[L](β)E[L](−β)=G(c+βs)G(c−βs)>1,\frac{\mathbb{E}[L](\beta)}{\mathbb{E}[L](-\beta)}=\frac{G\left(\frac{c+\beta}{s}\right)}{G\left(\frac{c-\beta}{s}\right)}>1,

and the ratio tends to ∞\infty as β/s→∞\beta/s \to\infty.

Proof. (i) By Lemma 4.1, Dt−1D_{t-1} is uniform and independent of ηt\eta_t, so rt=βDt−1+ηtr_t=\beta D_{t-1}+\eta_t is an equal mixture of ηt+β\eta_t+\beta and ηt−β\eta_t-\beta. Replacing β\beta by −β-\beta swaps the components. The variance follows because Dt−1D_{t-1} and ηt\eta_t are independent with mean 00, and FF is the mixture’s distribution function at −c-c.

(ii) If St=1S_t=1, then rt<−c≤0r_t<-c\le0, so Dt=−1D_t=-1. Hence P(St+1=1∣Ft,St=1)=P(−β+ηt+1<−c)=G(β−cs)\mathbb{P}(S_{t+1}=1\mid\mathcal{F}_t,S_t=1)=\mathbb{P}(-\beta+\eta_{t+1}<-c)=G\left(\frac{\beta-c}{s}\right), whatever the history. Each period of an episode continues it independently with this probability, so LL is geometric with E[L]=1/(1−G((β−c)/s))=1/G((c−β)/s)\mathbb{E}[L]=1/(1-G((\beta-c)/s))=1/G((c-\beta)/s), which is strictly increasing in β\beta because GG is.

(iii) Apply (ii) at ±β\pm\beta. As β/s→∞\beta/s\to\infty, the numerator tends to 11 and the denominator to 00.

For Gaussian news and c=0c=0, E[L](β)=1/Φ(−β/s)\mathbb{E}[L](\beta)=1/\Phi(-\beta/s): β=s\beta=s gives expected episodes of 6.3 periods, against 1.19 for β=−s\beta=-s, with identical volatility and stress frequency (Figure 1a).

Corollary 4.3 (Marginal-based rules cannot separate feedback). Let Ψ\Psi be any functional of the one-period return law and the gross turnover: volatility, value-at-risk, expected shortfall, stress frequency at any threshold, kurtosis, or a volume cap. Then Ψ\Psi takes the same value on the herd-dominated market (h,α)=(m,0)(h,\alpha)=(m,0) and the arbitrage-dominated market (h,α)=(0,m)(h,\alpha)=(0,m), yet their expected stress-episode lengths differ by the factor in Theorem 4.2(iii), which is unbounded in λm/s\lambda m/s.

Proof. The two markets have β=±λm\beta=\pm\lambda m and the same gross turnover mm. Apply Theorem 4.2(i) and (iii).

A regulator must therefore look at serial structure. The simplest serial statistic already works.

Proposition 4.4 (Serial statistics see feedback). If, in addition, ξ1\xi_1 has finite variance, then E[rt]=0\mathbb{E}[r_t]=0 and Cov⁡(rt,rt−1)=βE∣rt−1∣\operatorname{Cov}(r_t,r_{t-1})=\beta\mathbb{E}|r_{t-1}|. Hence the first-order autocorrelation ρ1\rho_1 and the two-period variance ratio VR⁡(2)−1=ρ1\operatorname{VR}(2)-1=\rho_1 [15] have the sign of β\beta.

Proof. rtr_t is a symmetric mixture by Theorem 4.2(i), so E[rt]=0\mathbb{E}[r_t]=0. Since ηt\eta_t is independent of rt−1r_{t-1} with mean 00, E[rtrt−1]=βE[Dt−1rt−1]=βE∣rt−1∣\mathbb{E}[r_tr_{t-1}]=\beta\mathbb{E}[D_{t-1}r_{t-1}]=\beta\mathbb{E}|r_{t-1}|, and E∣rt−1∣>0\mathbb{E}|r_{t-1}|>0 because rt−1r_{t-1} has a continuous law. For a stationary series, VR⁡(2)=Var⁡(rt+rt−1)/(2Var⁡(rt))=1+ρ1\operatorname{VR}(2)=\operatorname{Var}(r_t+r_{t-1})/(2\operatorname{Var}(r_t))=1+\rho_1.

The next theorem gives the exact cost of reading the sign.

Theorem 4.5 (Tight sample complexity of detecting feedback). Assume Gaussian news (G=ΦG=\Phi) and 0<∣β∣≤s0<|\beta|\le s. Given r0,…,rnr_0,\ldots,r_n, the test that outputs “herding” iff q^=1n∑t=1n1[Dt=Dt−1]>12\hat q=\frac{1}{n}\sum_{t=1}^{n}\mathbf{1}[D_t=D_{t-1}]>\frac{1}{2} identifies sgn⁡(β)\operatorname{sgn}(\beta) with error at most δ\delta whenever

n≥9(sβ)2log⁡1δ.n\ge9\left(\frac{s}{\beta}\right)^2\log\frac{1}{\delta}.

Conversely, for δ<1/4\delta<1/4, every test that observes the full returns has error larger than δ\delta under β\beta or under −β-\beta unless

n≥12(sβ)2log⁡14δ.n\ge\frac{1}{2}\left(\frac{s}{\beta}\right)^2\log\frac{1}{4\delta}.

Proof. Upper bound. By Lemma 4.1, nq^∼Bin⁡(n,q)n\hat{q} \sim\operatorname{Bin}(n,q) with q=Φ(β/s)q = \Phi(\beta/s). Say β>0\beta> 0; the other case is symmetric. Hoeffding’s inequality [11] gives P(q^≤12)≤exp⁡(−2n(q−12)2)\mathbb{P}(\hat{q} \le\frac{1}{2}) \le\exp(-2n(q-\frac{1}{2})^2). For 0<x≤10 < x \le1, Φ(x)−12=∫0xφ≥xφ(1)\Phi(x)-\frac{1}{2} = \int_{0}^{x}\varphi\ge x\varphi(1), and 2φ(1)2>1/92\varphi(1)^2 > 1/9. Hence the error is at most exp⁡(−nβ2/(9s2))≤δ\exp(-n\beta^2/(9s^2)) \le\delta.

Lower bound. Let P±P_{\pm} be the laws of (r0,…,rn)(r_0,\ldots,r_n) under ±β\pm\beta. The law of r0r_0 is the same under both (Theorem 4.2(i)). Conditional on the past, rt∼N(±βDt−1,s2)r_t \sim\mathcal{N}(\pm\beta D_{t-1},s^2): two normals whose means differ by 2∣β∣2|\beta|, so each conditional KL divergence equals 2β2/s22\beta^2/s^2, and by the chain rule KL⁡(P+∥P−)=2nβ2/s2\operatorname{KL}(P_+\|P_-) = 2n\beta^2/s^2. The Bretagnolle–Huber inequality [19] gives, for any test ψ\psi, P+(ψ≠+)+P−(ψ≠−)≥12e−2nβ2/s2P_+(\psi\ne+)+P_-(\psi\ne-) \ge\frac{1}{2}e^{-2n\beta^2/s^2}. So the larger of the two errors is at least 14e−2nβ2/s2\frac{1}{4}e^{-2n\beta^2/s^2}, which exceeds δ\delta unless n≥s22β2log⁡14δn \ge\frac{s^2}{2\beta^2}\log\frac{1}{4\delta}.

The sign-agreement test reads only DtD_t, yet it matches the full-information lower bound up to a constant. This is the theoretical basis for using learned policies as measurement devices: a policy whose action depends on sgn⁡(mom⁡t)\operatorname{sgn}(\operatorname{mom}_t) is, in effect, this test.

Proposition 4.6 (A herd-follower’s own stress is short). Consider a herd-follower with one-period momentum, exposure et∈{0,Δ,…,1}e_t \in\{0,\Delta,\ldots,1\}, step Δ\Delta, net return gt+1=etRt+1−ctx∣et−et−1∣g_{t+1}=e_tR_{t+1}-c_{\mathrm{tx}}|e_t-e_{t-1}|, and strategy stress 1[gt+1<−θ]\mathbf{1}[g_{t+1}<-\theta] with θ>ctxΔ\theta>c_{\mathrm{tx}}\Delta. On every return path, its stress episodes last at most 1/Δ1/\Delta periods. With Δ=12\Delta=\frac{1}{2}, that is two periods.

Proof. Strategy stress at t+1t+1 requires etRt+1<−θ+ctxΔ<0e_tR_{t+1}<-\theta+c_{\mathrm{tx}}\Delta<0, so et>0e_t>0 and Rt+1<0R_{t+1}<0, and the agent then sells: et+1=et−Δe_{t+1}=e_t-\Delta. Exposure therefore falls by Δ\Delta in each stressed period, and stress is impossible at zero exposure, where g≥−ctxΔ>−θg\ge-c_{\mathrm{tx}}\Delta>-\theta.

Proposition 4.6 and Theorem 4.2(ii) together explain a pattern found empirically in [2] (§ 7). At the strategy level, herd-followers show frequent but brief stress. At the market level, more herding lengthens stress episodes, because E[L]\mathbb{E}[L] increases in hh. Strategy-level measurements capture the former; regulation should care about the latter.

Remark 4.7 (Exact constant-product impact). Let each population trade a net m<1m<1 units of the yy-reserve per period. With the exact impact 2log⁡(1+N)2\log(1+N), a buy shifts the return by β+=2log⁡(1+m)\beta_+=2\log(1+m) and a sell by −β−-\beta_-, where β−=−2log⁡(1−m)>β+\beta_-=-2\log(1-m)>\beta_+. The sign chain stays Markov, and the proof of Theorem 4.2(ii) goes through state by state. In the herd market the continuation probability of a stress episode is G(β−−cs)G\left(\frac{\beta_- - c}{s}\right); in the arbitrage market it is G(−β+−cs)G\left(\frac{-\beta_+ - c}{s}\right). The ratio in part (iii) becomes G(c+β+s)G(c−β−s)\frac{G\left(\frac{c+\beta_+}{s}\right)}{G\left(\frac{c-\beta_-}{s}\right)}, which is larger still. Part (i) and Corollary 4.3 hold up to O(m2)O(m^2) terms.

Agents playing together

Section 4 fixed the mass of each feedback type. We now let profit-seeking agents choose their behavior in a pool where herd-followers already trade, and ask what learning converges to.

Players. A mass h>0h>0 of herd-followers trades hDt−1hD_{t-1} every period; by Lemma 3.2, a large enough momentum bonus makes this their behavior whatever it earns. A continuum of nonatomic rational traders with total mass Xˉ\bar{X} each choose one of three rules: ride (MM), opening a one-unit position in direction Dt−1D_{t-1}; lean (CC), opening a one-unit position in direction −Dt−1-D_{t-1}; or stay out (00). Positions opened in batch tt are closed in batch t+1t+1, and each round trip pays a fee κ>0\kappa>0. Write xM,xCx_M,x_C for the masses choosing MM and CC, and w=xM−xCw=x_M-x_C.

Prices. The batch-tt net flow is the herd flow plus the rational traders’ openings minus their closings, with impact

It=λhDt−1+λw(Dt−1−Dt−2),rt=It+ηt.I_t = \lambda h D_{t-1} + \lambda w\left(D_{t-1} - D_{t-2}\right), \qquad r_t = I_t + \eta_t.

Orders within a batch execute in uniformly random order against linear impact, so each order’s expected fill is the batch midpoint Pˉt=log⁡pt−1+It/2\bar{P}_t = \log p_{t-1} + I_t/2, in the spirit of batch auctions [5]. A rider earns Dt−1(Pˉt+1−Pˉt)D_{t-1}(\bar{P}_{t+1} - \bar{P}_t), and a leaner earns its negative.

The sign process. Call period tt continuing if Dt−1=Dt−2D_{t-1} = D_{t-2} and fresh otherwise. The sign persists with probability

a=G(λhs) in continuing periods,b=G(λ(h+2w)s) in fresh periods.a = G\left(\frac{\lambda h}{s}\right) \text{ in continuing periods}, \qquad b = G\left(\frac{\lambda(h+2w)}{s}\right) \text{ in fresh periods}.

because the rational openings and closings add 2λwDt−12\lambda wD_{t-1} only right after a flip. Whether a period is fresh is therefore a two-state Markov chain, and the stationary probability that the sign persists is

q(h,w)=b1−a+b(4)q(h,w) = \frac{b}{1-a+b} \tag*{(4)}

which is strictly increasing in ww and reduces to q=aq = a, the stay probability of Lemma 4.1, when w=0w = 0.

Lemma 5.1 (Own impact cancels). In the stationary regime, the expected per-period payoffs are

uM=λhq(h,w)−κ,uC=−λhq(h,w)−κ,u0=0.u_M = \lambda h q(h,w) - \kappa, \qquad u_C = -\lambda h q(h,w) - \kappa, \qquad u_0 = 0.

They depend on the rational traders’ own flow only through qq.

Proof. Pˉt+1−Pˉt=It/2+ηt+It+1/2\bar{P}_{t+1} - \bar{P}_t = I_t/2 + \eta_t + I_{t+1}/2. Let ρ=E[DtDt−1]=2q−1\rho= \mathbb{E}[D_tD_{t-1}] = 2q - 1; by stationarity it is also E[Dt−1Dt−2]\mathbb{E}[D_{t-1}D_{t-2}]. Then

E[Dt−1It]=λh+λw(1−ρ),E[Dt−1It+1]=λhρ+λw(ρ−1),\mathbb{E}[D_{t-1}I_t] = \lambda h + \lambda w(1-\rho), \qquad\mathbb{E}[D_{t-1}I_{t+1}] = \lambda h\rho+ \lambda w(\rho-1),

and E[Dt−1ηt]=0\mathbb{E}[D_{t-1}\eta_t] = 0 because ηt\eta_t is independent of Ft−1\mathcal{F}_{t-1}. Averaging the two, the ww terms cancel, and the expected gross payoff from riding is λh(1+ρ)/2=λhq\lambda h(1+\rho)/2 = \lambda hq. Leaning is the opposite position. Subtracting the fee gives both formulas.

The ww terms cancel because a round trip on a path-independent pool cannot profit from its own impact. A rational trader profits only by trading ahead of predictable herd flow.

Theorem 5.2 (Rational agents join the herd; learning tips). Let uM(x)=λhq(h,x)−κu_M(x) = \lambda hq(h,x) - \kappa with xC=0x_C = 0, and suppose h>0h > 0.

(i) No equilibrium leans. Leaning is strictly dominated, uC<0=u0u_C < 0 = u_0, so xC=0x_C = 0 in every Nash equilibrium. Without an external price anchor, negative feedback is never a rational response to herd flow.

(ii) Strategic complements. uM(x)u_M(x) is strictly increasing in xx. Hence x=0x = 0 is an equilibrium iff λhG(λh/s)≤κ\lambda hG(\lambda h/s) \le\kappa; x=Xˉx = \bar{X} is an equilibrium iff λhq(h,Xˉ)≥κ\lambda hq(h,\bar{X}) \ge\kappa; and at most one interior equilibrium x∗x^* exists, solving uM(x∗)=0u_M(x^*) = 0, and it is unstable. In particular, for κ∈(λhG(λhs),λhq(h,Xˉ))\kappa\in\left(\lambda hG\left(\frac{\lambda h}{s}\right),\lambda hq(h,\bar{X})\right) both 00 and Xˉ\bar{X} are equilibria.

(iii) Learning tips. Consider any sign-preserving learning dynamics, with x˙M≥0\dot{x}_M \ge0 when uM≥0u_M \ge0 and x˙C<0\dot{x}_C < 0 when xC>0x_C > 0, at speed bounded away from zero wherever the payoff is. This class includes replicator dynamics and expected softmax policy gradient [18]. Then xC→0x_C \to0. If λhq(h,xM(0)−xC(0))>κ\lambda hq(h,x_M(0)-x_C(0))>\kappa, then xM→Xˉx_M \to\bar{X}; if λhq(h,xM(0))<κ\lambda hq(h,x_M(0))<\kappa, then xM→0x_M \to0. With xC(0)=0x_C(0)=0, the basin boundary is x∗x^*.

(iv) Joining amplifies stress. At threshold c=0c=0, down-runs have expected length

E[L](x)=1+G(λ(h+2x)/s)G(−λh/s),\mathbb{E}[L](x)=1+\frac{G(\lambda(h+2x)/s)}{G(-\lambda h/s)},

which is strictly increasing in xx. The high-participation equilibrium therefore has strictly more persistent stress than herding alone.

Proof. (i) From Lemma 5.1, uC=−λhq−κ<0u_C=-\lambda hq-\kappa<0 because q>0q>0 and κ>0\kappa>0.

(ii) bb is strictly increasing in xx because GG is, and ∂q/∂b=(1−a)/(1−a+b)2>0\partial q/\partial b=(1-a)/(1-a+b)^2>0, so uMu_M is strictly increasing. At x=0x=0, q=aq=a. The three cases are the equilibrium conditions for an increasing payoff against the outside option 0. An interior root is unstable because uMu_M is negative below it and positive above it.

(iii) Since uC<0u_C<0 everywhere, xCx_C decreases to 0. Write w(t)=xM(t)−xC(t)w(t)=x_M(t)-x_C(t). In the first case uM(w(0))>0u_M(w(0))>0, so xMx_M rises while xCx_C falls; then ww rises, and since uMu_M is increasing it stays positive, so xM→Xˉx_M\to\bar{X}. In the second case uM(w(0))≤uM(xM(0))<0u_M(w(0))\le u_M(x_M(0))<0, so xMx_M falls; then w(t)≤xM(t)≤xM(0)w(t)\le x_M(t)\le x_M(0) for all tt, uMu_M stays negative, and xM→0x_M\to0.

(iv) At c=0c=0, a down-run starts with a flip, so its first continuation is a fresh period, which persists with probability bb; every later continuation is a continuing period, which persists with probability aa. Hence E[L]=1+b/(1−a)\mathbb{E}[L]=1+b/(1-a) with 1−a=G(−λh/s)1-a=G(-\lambda h/s), and bb is strictly increasing in xx.

Corollary 5.3 (What a uniform fee can and cannot do). A uniform round-trip fee κ>λhq(h,Xˉ)\kappa>\lambda hq(h,\bar{X}) removes every equilibrium with rational participation. No uniform fee induces leaning: by part (i), uC<0u_C<0 for every κ≥0\kappa\ge0. Negative feedback against herds must come from outside this game, for example from arbitrage to an external reference price or from agents who are paid to take the other side.

Proof. For κ>λhq(h,Xˉ)≥λhq(h,x)\kappa>\lambda hq(h,\bar{X})\ge\lambda hq(h,x), uM(x)<0u_M(x)<0 for every xx, so xM=0x_M=0 in equilibrium. The second claim restates part (i).

Theorem 5.2 connects the archetypes. The return-maximizing agent, placed in a pool with herds, learns to trade ahead of herd flow, which is the behavior the manipulator archetype was designed to proxy. How a profit-seeking agent behaves depends on who else is in the pool, so the archetypes cannot be studied in isolation. With λ=s=1\lambda=s=1, h=0.3h=0.3, κ=0.2\kappa=0.2, and Xˉ=1\bar{X}=1, the basin boundary is x∗≈0.21x^*\approx0.21; softmax policy-gradient learners starting from 15% riders converge to 0.9%, and from 40% to 98%, while down-runs lengthen from 2.62 periods (herds alone) to 3.59 (Appendix A).

Adding profit-maximizing market makers

Corollary 5.3 states that negative feedback must come from agents who are paid to take the other side. We now add them: competitive market makers (MMs) who supply liquidity next to the pool and earn a spread for absorbing flow, as wholesalers internalize retail orders. An MM differs from a leaning trader in one respect only: it is paid to trade rather than paying to trade.

Market makers. A competitive mass of MMs absorbs a share σ∈[0,σˉ]\sigma\in[0,\bar{\sigma}] of herd flow in batch tt, at the batch midpoint, and earns a net spread ϕ>0\phi> 0 per unit absorbed. MMs carry the inventory −σhDt−1-\sigma hD_{t-1} for one period and unwind it in batch t+1t+1. Riders are as in Section 5, with mass w=xMw=x_M (leaning remains dominated for them) and fee κ\kappa.

Prices. The herd’s net impact in batch tt is (1−σ)hDt−1(1-\sigma)hD_{t-1}, plus the MMs’ unwinding σhDt−2\sigma hD_{t-2}. Adding the riders’ round trips,

It=λhDt−1+λz(Dt−1−Dt−2),z=w−σh.I_t=\lambda hD_{t-1}+\lambda z\left(D_{t-1}-D_{t-2}\right),\qquad z=w-\sigma h.

So MMs enter exactly as negative riders, and everything depends on the net ignition kick zz: rider capital minus the herd flow absorbed by MMs. The fresh-period stay probability becomes b(z)=G(λ(h+2z)/s)b(z)=G(\lambda(h+2z)/s), the continuing-period probability a=G(λh/s)a=G(\lambda h/s) is unchanged, and q(z)=b(z)/(1−a+b(z))q(z)=b(z)/(1-a+b(z)).

Theorem 6.1 (Market makers). In the stationary regime:

(i) Payoffs. Riders earn uM=λhq(z)−κu_M=\lambda hq(z)-\kappa per unit; MMs earn uMM=ϕ−λhq(z)u_{MM}=\phi-\lambda hq(z) per unit absorbed. Own impact cancels for both groups.

(ii) Persistence depends only on zz. At c=0c=0, E[L]=1+b(z)/G(−λh/s)\mathbb{E}[L]=1+b(z)/G(-\lambda h/s), strictly increasing in zz. MMs shorten stress, and when σh>w\sigma h>w stress is shorter than with herds alone. MMs dampen the start of a trend but not an established one, since aa is unaffected.

(iii) Complements within groups, substitutes across groups. uMu_M is increasing in ww and decreasing in σ\sigma; uMMu_{MM} is increasing in σ\sigma and decreasing in ww.

(iv) Corner equilibria. (Xˉ,σˉ)(\bar{X},\bar{\sigma}) is an equilibrium iff κ≤λhq(Xˉ−σˉh)≤ϕ\kappa\leq\lambda hq(\bar{X}-\bar{\sigma}h)\leq\phi; (Xˉ,0)(\bar{X},0) iff λhq(Xˉ)≥max⁡(κ,ϕ)\lambda hq(\bar{X})\geq\max(\kappa,\phi); (0,σˉ)(0,\bar{\sigma}) iff λhq(−σˉh)≤min⁡(κ,ϕ)\lambda hq(-\bar{\sigma}h)\leq\min(\kappa,\phi); and (0,0)(0,0) iff ϕ≤λhG(λh/s)≤κ\phi\leq\lambda hG(\lambda h/s)\leq\kappa. An equilibrium with interior ww requires λhq(z)=κ\lambda hq(z)=\kappa, and one with interior σ\sigma requires λhq(z)=ϕ\lambda hq(z)=\phi; both are unstable within their own group under sign-preserving dynamics.

*(v) Liquidity collapse and hysteresis. Let κ<ϕ\kappa<\phi and λhq(−σˉh)>κ\lambda hq(-\bar{\sigma}h)>\kappa, so riders are always active, and define Xˉc\bar{X}_c and Xˉlo\bar{X}_{lo} by

λhq(Xˉc−σˉh)=ϕ,λhq(Xˉlo)=ϕ,\lambda hq(\bar{X}_c-\bar{\sigma}h)=\phi,\qquad\lambda hq(\bar{X}_{lo})=\phi,

assuming both exist. Since qq is strictly increasing, Xˉlo=Xˉc−σˉh\bar{X}_{lo}=\bar{X}_c-\bar{\sigma}h: the hysteresis band is exactly as wide as the MMs’ absorption capacity. For Xˉ<Xˉc\bar{X}<\bar{X}_c the full-liquidity equilibrium (Xˉ,σˉ)(\bar{X},\bar{\sigma}) exists and is stable. At Xˉc\bar{X}_c it disappears, MMs exit, and E[L]\mathbb{E}[L] jumps upward by

ΔL=G(λ(h+2Xˉc)/s)−G(λ(h+2Xˉc−2σˉh)/s)G(−λh/s)>0.\Delta_L=\frac{G(\lambda(h+2\bar{X}_c)/s)-G(\lambda(h+2\bar{X}_c-2\bar{\sigma}h)/s)}{G(-\lambda h/s)}>0.

For Xˉ∈(Xˉlo,Xˉc)\bar{X}\in(\bar{X}_{lo},\bar{X}_c) both (Xˉ,σˉ)(\bar{X},\bar{\sigma}) and (Xˉ,0)(\bar{X},0) are stable, so once liquidity is lost it returns only when rider capital falls below Xˉlo\bar{X}_{lo}.

Proof. (i) Riders follow the proof of Lemma 5.1 with ww replaced by zz; the zz terms cancel. An MM holds −σhDt−1-\sigma hD_{t-1} from midpoint tt to midpoint t+1t+1, and the same computation gives gross payoff −λhq-\lambda hq per unit, to which the spread ϕ\phi is added.

(ii) This is the proof of Theorem 5.2(iv) with b=b(z)b=b(z).

(iii) qq is strictly increasing in z=w−σhz=w-\sigma h.

(iv) Each condition says that every active group earns at least 0 and every inactive group at most 0. At an interior point, a group’s payoff is increasing in its own share, so a small deviation grows.

(v) For Xˉ<Xˉc\bar{X} < \bar{X}_{c}, λhq(Xˉ−σˉh)<ϕ\lambda h q(\bar{X} - \bar{\sigma}h) < \phi, and by assumption it exceeds κ\kappa, so both groups earn strictly positive payoffs at the corner, which is therefore stable. For Xˉ>Xˉc\bar{X} > \bar{X}_{c}, uMM(Xˉ,σ)≤uMM(Xˉ,σˉ)<0u_{MM}(\bar{X},\sigma) \leq u_{MM}(\bar{X},\bar{\sigma}) < 0 for every σ\sigma, since uMMu_{MM} is increasing in σ\sigma; so MMs exit, and zz jumps from Xˉc−σˉh\bar{X}_{c} - \bar{\sigma}h to Xˉc\bar{X}_{c}, and ΔL\Delta_L is the difference in (ii). For Xˉ>Xˉlo\bar{X} > \bar{X}_{lo}, uMM(Xˉ,0)=ϕ−λhq(Xˉ)<0u_{MM}(\bar{X},0) = \phi- \lambda h q(\bar{X}) < 0, so (Xˉ,0)(\bar{X},0) is a strict, stable equilibrium. On (Xˉlo,Xˉc)(\bar{X}_{lo},\bar{X}_{c}) both corners are strict.

Corollary 6.2 (Maker–taker pricing). Corollary 5.3 said no uniform fee induces leaning. With market makers, a direction-blind maker rebate that raises ϕ\phi does. Raising ϕ\phi raises Xˉc\bar{X}_{c} and Xˉlo\bar{X}_{lo}, which widens the range of rider capital over which full liquidity is an equilibrium and can be recovered after a collapse. A taker fee κ>λhq(Xˉ−σˉh)\kappa> \lambda h q(\bar{X} - \bar{\sigma}h) removes riders, which sets z=−σˉhz = -\bar{\sigma}h and minimizes E[L]\mathbb{E}[L].

Proof. qq is strictly increasing, so Xˉc\bar{X}_{c} and Xˉlo\bar{X}_{lo} solve q(⋅)=ϕ/(λh)q(\cdot) = \phi/(\lambda h) at points that increase with ϕ\phi. The taker-fee claim follows from (iv) and (ii).

What changes with market makers. Stabilizing behavior now arises in equilibrium: leaning is still dominated for traders who pay to trade, but profitable for MMs whenever ϕ>λhq\phi> \lambda hq, so what separates a stabilizer from a non-participant is who pays the spread. Persistence is set by one index, z=w−σhz = w - \sigma h, so each unit of absorption offsets one unit of rider capital, and MMs alone (z=−0.3z = -0.3 in the example of Figure 1b) cut expected down-runs from 2.62 to 2.00 periods. The market becomes fragile: tipping now happens in liquidity, with a discontinuous jump in stress persistence and hysteresis, as in liquidity spirals [4] and the Flash Crash [12]. And the effective instrument is a maker–taker schedule, which is behavior-specific, rather than a uniform fee.

Connection to empirical evidence

The companion study [2] trains the three archetypes of § 2 with DQN and REINFORCE on hourly bars for AAVE, ETC, and SUI from March 2024 to February 2026, as price-takers in separate environments, and attributes the learned policies with SHAP. Each policy relies on the signals its objective implies: momentum for the herd-follower, drawdown and cross-asset returns for the rational agent, and short-term volatility and position for the manipulator. Three of its findings bear on the theory.

Herders: brief own stress, longer market stress. The study finds that herd-follower strategies have the most frequent but shortest stress episodes. This is Proposition 4.6: with exposure steps of 12\frac{1}{2}, a herd-follower’s own episodes last at most two periods on every path. Theorem 4.2(ii) shows that the same agents lengthen market stress, an effect that strategy-level metrics cannot detect.

Rational agents: stabilizing alone, not together. The study’s rational agents, trading alone, show the most stable risk-adjusted returns, and its manipulators the rarest but most persistent stress. Theorem 5.2 shows that a return-maximizing agent sharing a pool with herds never leans against them and, above a computable participation threshold, learns to ride them, which is the front-running behavior the manipulator archetype proxies, and which lengthens stress. Reading rational agents as stabilizers is therefore valid only where something else supplies negative feedback, such as market makers (Theorem 6.1) or arbitrage to an external price.

What market-level data show. In the study’s data, the two-hour variance ratio is 1.010–1.017, statistically indistinguishable from one. Therefore, by Proposition 4.4, the full sample shows no net feedback at the hourly horizon. Stress nonetheless clusters: runs of hours with returns below −2% average 1.09–1.13 hours, against 1.02–1.04 under independence, and absolute returns have autocorrelation 0.22–0.27. This persistence comes from volatility clustering, which Assumption 1 excludes. Empirical tests should therefore apply the serial statistics of § 4 conditionally, relating variation in sign persistence to archetype intensity after controlling for volatility rather than to full-sample levels. The theory yields three such predictions: herd intensity raises market stress persistence without raising stress frequency at fixed gross flow (Theorem 4.2); the sign of VR⁡(2)−1\operatorname{VR}(2)-1 tracks the net index zz (Proposition 4.4, Theorem 6.1); and market-maker participation falls discontinuously as trend-chasing capital rises and recovers only after a larger decline (Theorem 6.1(v)).

Implications for regulation

Corollary 4.3 shows that limits on volatility, value-at-risk, loss frequency, or turnover cannot distinguish stabilizing from destabilizing populations of agents. Theorem 4.5 shows that the distinction is cheap to read from serial statistics at a rate no method can beat. Corollaries 5.3 and 6.2 show that a uniform fee can deter riders but never induce stabilizing trades, while a maker rebate does and widens the range of trend-chasing capital over which liquidity survives and can be restored. Instruments should act on feedback rather than on one-period statistics.

Conclusion

Defining trading agents by their objectives traces a path from rewards to order flow to market stress. On that path, the one-period statistics regulators watch are provably blind, while serial statistics separate stabilizing from destabilizing feedback at an optimal rate. When agents trade together, profit-seeking agents amplify herds, and market makers that are paid to absorb flow restore stability at the cost of fragile, hysteretic liquidity. Instruments that act on feedback, such as maker–taker pricing, are the effective response.

Numerical verification

Every closed form in § 4–§ 6 was checked by Monte Carlo simulation with Gaussian news and λ=s=1\lambda=s=1. Table 1 reports simulated values with theoretical values in parentheses.

Result and settingSimulated (theory)
Thm. 4.2, β=+0.8\beta=+0.8, c=1c=1, 2×1062 \times10^{6} periodsVar⁡(r)\operatorname{Var}(r) 1.6398 (1.6400); FF 0.2282 (0.2283); E[L]\mathbb{E}[L] 1.7247 (1.7263); qq 0.7883 (0.7881)
Thm. 4.2, β=−0.8\beta=-0.8, c=1c=1Var⁡(r)\operatorname{Var}(r) 1.6398 (1.6400); FF 0.2287 (0.2283); E[L]\mathbb{E}[L] 1.0372 (1.0373); qq 0.2117 (0.2119)
Prop. 4.6, Δ=12\Delta=\frac{1}{2}, 3×1053 \times10^{5} periodslongest own-stress episode 2 (≤2\le2); mean 1.08
Lemma 5.1, h=0.3h=0.3, (xM,xC)=(0,0)(x_M,x_C)=(0,0), 10610^{6} periodsrider gross 0.1847 (0.1854); qq 0.6181 (0.6179); E[L]\mathbb{E}[L] 2.6164 (2.6172)
(xM,xC)=(0.5,0)(x_M,x_C)=(0.5,0)0.2102 (0.2108); 0.7026 (0.7027); 3.3608 (3.3638)
(xM,xC)=(1,0)(x_M,x_C)=(1,0)0.2155 (0.2164); 0.7215 (0.7214); 3.5860 (3.5891)
(xM,xC)=(0,0.5)(x_M,x_C)=(0,0.5)0.1173 (0.1163); 0.3874 (0.3877); 1.6295 (1.6333)
Thm. 5.2(iii), κ=0.2\kappa=0.2, Xˉ=1\bar{X}=1, softmax policy gradientfrom shares (0.15,0.15,0.70)(0.15,0.15,0.70): (0.009,0.001,0.991)(0.009,0.001,0.991); from (0.40,0.15,0.45)(0.40,0.15,0.45): (0.981,0.001,0.018)(0.981,0.001,0.018); basin boundary x∗≈0.21x^{*}\approx0.21
Thm. 6.1(i)–(ii), (w,σ)=(0,0.5)(w,\sigma)=(0,0.5), z=−0.15z=-0.15rider gross 0.1699 (0.1701); MM gross −0.1699-0.1699 (−0.1701-0.1701); E[L]\mathbb{E}[L] 2.308 (2.309)
(w,σ)=(0.5,1)(w,\sigma)=(0.5,1), z=0.20z=0.200.1988 (0.1995); −0.1988-0.1988 (−0.1995-0.1995); 2.996 (2.984)
(w,σ)=(1,0.5)(w,\sigma)=(1,0.5), z=0.85z=0.850.2168 (0.2157); −0.2168-0.2168 (−0.2157-0.2157); 3.571 (3.558)
Thm. 6.1(v), κ=0.12\kappa=0.12, ϕ=0.2\phi=0.2, σˉ=1\bar{\sigma}=1Xˉc=0.510\bar{X}_c=0.510, Xˉlo=0.210\bar{X}_{lo}=0.210; E[L]\mathbb{E}[L] jumps 3.000→3.3733.000 \to3.373; learning dynamics lose MMs at Xˉ≈0.6\bar{X}\approx0.6 on the way up and recover them at Xˉ≈0.1\bar{X}\approx0.1 on the way down

Table 1. Simulated versus theoretical values. Payoffs are gross of fees, per unit per period.

Simulation design. § 4 simulates (3) directly. § 5–§ 6 simulate the batch model with impact ItI_t, fill each order at the batch midpoint, and compute payoffs on round trips from batch tt to t+1t+1. For Theorem 5.2(iii), each of 400 iterations estimates the three payoffs from a fresh simulation of 2×1042 \times10^{4} periods at the current shares and applies the expected softmax policy-gradient update with step size 4. For Theorem 6.1(v), rider and MM shares follow logistic sign-preserving dynamics on the closed-form payoffs, run to convergence at each Xˉ\bar{X} while Xˉ\bar{X} is swept up from 0.05 to 0.8 and back. Learning slows near each threshold, so the simulated collapse and recovery points lie outside the band [Xˉlo,Xˉc][\bar{X}_{lo},\bar{X}_c] rather than on its edges, as stability predicts. The code is provided as supplementary material.

References

  1. [1]Hayden Adams, Noah Zinsmeister, and Dan Robinson. Uniswap v2 core. Whitepaper, 2020.
  2. [2]Irene Aldridge, Lourenço Belleau-Bowhnik, James Chan, Josh Chang, Kelly Deng, Pranav Dhingra, Xinda Ding, Haodong Fang, Satya Hanumanchi, Carter He, Shuang He, Yadong Hou, Vallab Karanam, Rithvik Kasarla, Tishya Khanna, Jihyun Kim, Sarah Konety, Abhinav Kotta, Anderson Lo, Jake Luther, Omar Mian, Ishan Patwardhan, Matteo Perona, Chenyue Qi, Shashank Ramachandran, Shubham Rateria, Jake Swain, David Thai, Michael Tversky, Susan Wu, Dean Xu, Helenna Yin, and Jiachun Zheng. Agentic AI in automated markets. arXiv:2610.22522, 2026.arxiv.org/abs/2610.22522
  3. [3]Guillermo Angeris and Tarun Chitra. Improved price oracles: Constant function market makers. In Proceedings of the 2nd ACM Conference on Advances in Financial Technologies (AFT ’20), pages 80–91. ACM, 2020.DOI
  4. [4]Markus K. Brunnermeier and Lasse Heje Pedersen. Market liquidity and funding liquidity. Review of Financial Studies, 22(6):2201–2238, 2009.
  5. [5]Eric Budish, Peter Cramton, and John Shim. The high-frequency trading arms race: Frequent batch auctions as a market design response. Quarterly Journal of Economics, 130(4):1547–1621, 2015.
  6. [6]Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello. Artificial intelligence, algorithmic pricing, and collusion. American Economic Review, 110(10):3267–3297, 2020.DOI
  7. [7]Philip Daian, Steven Goldfeder, Tyler Kell, Yunqi Li, Xueyuan Zhao, Iddo Bentov, Lorenz Breidenbach, and Ari Juels. Flash boys 2.0: Frontrunning in decentralized exchanges, miner extractable value, and consensus instability. In Proceedings of the 2020 IEEE Symposium on Security and Privacy, pages 910–927, 2020.DOI
  8. [8]J. Bradford De Long, Andrei Shleifer, Lawrence H. Summers, and Robert J. Waldmann. Positive feedback investment strategies and destabilizing rational speculation. Journal of Finance, 45(2):379–395, 1990.
  9. [9]Lawrence R. Glosten and Paul R. Milgrom. Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of Financial Economics, 14(1):71–100, 1985.DOI
  10. [10]Sanford J. Grossman and Merton H. Miller. Liquidity and market structure. Journal of Finance, 43(3):617–633, 1988.
  11. [11]Wassily Hoeffding. Probability inequalities for sums of bounded random variables. Journal of the American Statistical Association, 58(301):13–30, 1963.
  12. [12]Andrei Kirilenko, Albert S. Kyle, Mehrdad Samadi, and Tugkan Tuzun. The flash crash: High-frequency trading in an electronic market. Journal of Finance, 72(3):967–998, 2017.
  13. [13]Albert S. Kyle. Continuous auctions and insider trading. Econometrica, 53(6):1315–1335, 1985.DOI
  14. [14]Blake LeBaron. Agent-based computational finance. In Leigh Tesfatsion and Kenneth L. Judd, editors, Handbook of Computational Economics, volume 2, pages 1187–1233. Elsevier, 2006.
  15. [15]Andrew W. Lo and A. Craig MacKinlay. Stock market prices do not follow random walks: Evidence from a simple specification test. Review of Financial Studies, 1(1):41–66, 1988.
  16. [16]Paul Milgrom and John Roberts. Rationalizability, learning, and equilibrium in games with strategic complementarities. Econometrica, 58(6):1255–1277, 1990.DOI
  17. [17]Jason Milionis, Ciamac C. Moallemi, Tim Roughgarden, and Anthony Lee Zhang. Automated market making and loss-versus-rebalancing. arXiv:2208.06046, 2022.arxiv.org/abs/2208.06046
  18. [18]William H. Sandholm. Population Games and Evolutionary Dynamics. MIT Press, Cambridge, MA, 2010.
  19. [19]Alexandre B. Tsybakov. Introduction to Nonparametric Estimation. Springer Series in Statistics. Springer, New York, 2009.

Paper details

vlnc:W_pf7c6h9q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents