Reliability and Identifiability in Persona-Trained Monte Carlo: Variance Decomposition, Stability Bounds, and the Identifiability of Heterogeneous News Reaction
Source: arXiv:2607.04627 · Published 2026-07-06 · By Salavat Ishbulatov
TL;DR
This paper develops a rigorous statistical theory around Persona-Trained Monte Carlo (PTMC), a method that estimates distributions of market outcomes by simulating interacting limit-order-book trading bots whose behaviors are sampled from a learned persona distribution. The author decomposes the estimator variance into components attributable to persona heterogeneity and stochasticity within runs, deriving unbiased ANOVA-style variance estimators and an optimal allocation of simulation budget between persona draws and replications. A coupling-based stability bound quantifies how errors in estimating the persona distribution and policy affect the output distribution, yielding a three-part total reliability budget. The core technical contribution is an identification theory showing the conditions under which heterogeneous news sensitivity distributions can be recovered from aggregate market impact curves, exploiting a strict Jensen gap to detect heterogeneity and characterizing identifiability via odd moments and Hausdorff determinacy. Tests and consistent estimators for heterogeneity are developed, and formal separation theorems establish when PTMC outperforms homogeneous simulators and reduced-form forecasters. The results provide a careful error decomposition and structural guarantees rare in agent-based market simulation literature, though empirical validation and mean-field limits remain open.
Key findings
- The total variance of the PTMC estimator decomposes as Var(F) = σ2P + σ2w, where σ2P is variance across persona draws and σ2w is within-run variance, with unbiased ANOVA estimators available (Theorem 1).
- Optimal allocation of compute budget between persona draws D and replications R is R* = sqrt(c0σ2w / c1σ2P), minimizing variance to V* = (sqrt(c0σ2P) + sqrt(c1σ2w))^2 / B (Theorem 2).
- A coupling-based stability bound shows estimation error in the output scales linearly with population size K, horizon T, and errors in persona distribution (measured by Wasserstein-1 distance) and policy (total variation), with error ≤ 2M K T L W1(P, P̂) + 2M K T ε (Theorem 3).
- Under a uniform Doeblin minorization condition on the market state Markov chain, the horizon-dependent error can be bounded uniformly in T by replacing T with 1/α, α being the minorization constant (Theorem 4).
- Heterogeneous agent news sensitivity distribution Q is locally identifiable from the aggregate impact curve A(z) = E_Q[g(η z)] via odd moments and Hausdorff determinacy, with a strict Jensen gap detecting heterogeneity if Var_Q(η) > 0 (Theorem 5).
- If the response nonlinearity g is unknown, identifiability fails completely; heterogeneity cannot be distinguished from a homogeneous population with a deformed g (Proposition 1).
- Consistent estimators for mean and variance of η from event-study data are √n-consistent and asymptotically normal, and a boundary-corrected likelihood ratio test for homogeneity (Var(η)=0) obeys a 50:50 mixture of χ2 distributions on the boundary (Theorem 6 and 7).
- Separation theorems prove that homogeneous simulators have an irreducible Jensen bias error floor proportional to heterogeneity, while PTMC is consistent; observationally equivalent markets with different intervention responses create minimax lower bounds for reduced-form forecasters, embodying the Lucas critique (Theorems 8–10).
Threat model
The adversary is implicit in the form of the true underlying market data generation process, modeled as a limit order book with K heterogeneous trading agents each governed by individual personas drawn i.i.d. from a persona distribution P and controlled by a known policy. The adversary cannot alter the simulator mechanics or produce data outside this framework, but errors arise from uncertainty in estimating the persona distribution P and imperfect cloning of the trading policy π. Adversarial perturbation of input distributions or policies is bounded by Wasserstein and total variation distances, respectively, which measure how closely the simulated inputs match the true ones.
Methodology — deep read
The threat model assumes an adversary who may produce market data generated by limit order book interactions with agent heterogeneity according to some persona distribution P and policies. The learner attempts to estimate distribution functionals of market outcomes via Monte Carlo simulation.
The data consists of stochastic simulations with K neural policy bots, each instantiated with a persona sampled i.i.d. from a distribution P over a compact persona space. The policies πϕ map market states and persona parameters to action distributions, assumed Markov kernels with Lipschitz properties. The simulator runs for T steps yielding market states and order book trajectories. Functionals F of these trajectories are bounded measurable quantities of interest (e.g., price drawdowns).
The PTMC estimator involves two-stage nested sampling: D independent samples of persona populations ω (size K) are drawn, and for each persona population, R independent conditional simulations are run producing realizations of F, yielding estimates of the expectation µ = EP[F(X)].
Variance decomposition leverages the law of total variance: overall variance splits into between-persona variance σ2P and within-run variance σ2w. Unbiased ANOVA-style estimators for σ2P, σ2w are constructed based on the nested design with classic formulas. An optimal budget allocation derivation balances the cost of drawing new persona populations (expensive) vs replicating runs per persona.
Stability analysis uses coupling arguments, constructing joint distributions coupling two simulations with close persona distributions or close policies. The Lipschitz continuity assumptions on policies allow bounding total variation between distributions on simulated trajectories. The paper derives linear-in-KT bounds on estimation error as functions of Wasserstein distance between persona distributions and total variation distance between policies. By invoking ergodicity and Doeblin minorization conditions on the market state Markov chains, the horizon dependence can be controlled.
The identifiability theory for heterogeneous news reaction treats the persona-level news sensitivity η as a latent random coefficient with distribution Q. The aggregate market impact curve A(z) = EQ[g(η z)] is shown to be related to Q through a bounded saturating nonlinearity g. Using classical random-coefficient identification tools, the paper proves a strict Jensen gap that identifies heterogeneity, along with moment-based identification results relying on Hausdorff determinacy. Consistent estimators and boundary-corrected hypothesis tests are constructed from event-study data.
Separation theorems formally characterize when PTMC outperforms homogeneous simulators and reduced-form forecasters. They construct market pairs that are observationally indistinguishable yet differ in policy intervention effects, yielding minimax worst-case error lower bounds for forecasters relying on observational data alone, revealing a form of the Lucas critique.
No simulation experiments are included; the paper develops full mathematical proofs for the estimators, variance decompositions, coupling stability bounds, identifiability results, and separation theorems. The companion paper addresses empirical validation and framework specification.
Overall, the methodology is thoroughly theoretical: defining the stochastic system, specifying Markov kernels and assumptions, analytically decomposing variance, bounding perturbations via couplings, and proving statistical identification and inference properties in a functional-analytic framework. This rigorous mathematical approach dissects estimator reliability, error sources, and fundamental identifiability limits in agent-based market simulators.
Technical innovations
- Decomposition of PTMC estimator variance into persona-draw (between-population) and within-run components using unbiased ANOVA estimators, enabling optimal nested-simulation allocation.
- A coupling-based stability bound that quantifies how errors in the persona distribution (via Wasserstein distance) and in policy (via total variation) propagate to output estimand error, scaling linearly with population size and horizon length.
- An identifiability theory for heterogeneous news sensitivity distributions from aggregate impact curves, leveraging a strict Jensen gap and Hausdorff moment determinacy within a fixed nonlinear response family.
- Formal separation theorems proving irreducible bias floors for homogeneous simulators and minimax lower bounds for reduced-form observational forecasters, embodying the Lucas critique as a fundamental limit on intervention extrapolation.
Limitations
- The stability bounds scale linearly with population size K and horizon T, which can be pessimistic and large in practice.
- The uniform-in-horizon stability bound relies on a strong Doeblin minorization condition, whose validity for realistic limit order books remains an open question.
- No empirical or simulation experiments are reported; all results are theoretical and need companion numerical validation.
- Identification of heterogeneous news reaction requires knowledge of the nonlinear response function g; if g is unknown or misspecified, identifiability fails completely.
- Open problem remains on behavior and mean-field limits as population size K → ∞, limiting conclusions about scalability.
- The assumption that cloned policies have bounded total variation error may be hard to verify or guarantee in complex market environments.
Open questions / follow-ons
- How do the PTMC estimands and error decompositions behave in the large-population limit (K → ∞), and can mean-field results be established?
- Can the Doeblin minorization condition or an alternative ergodicity condition be verified or relaxed for realistic limit order book simulators?
- What are practical algorithms and procedures for estimating the persona distribution P and policy π with controlled total variation or Wasserstein errors in real market data settings?
- How do the identification and inference results extend when the nonlinear response function g is itself uncertain or adaptively learned?
Why it matters for bot defense
For bot-defense and CAPTCHA practitioners, this paper highlights the importance of rigorous variance decomposition and nested-simulation allocation in estimating quantities dependent on heterogeneous agent populations, emphasizing how persona heterogeneity contributes irreducible variance unaffected by replication alone. The coupling-based stability bounds provide a framework to quantify the impact of input model misestimation on final outputs, a crucial consideration for assessing robustness under distribution shifts. The identification theory for latent heterogeneity from aggregate functional responses may inspire methods to detect bot behavioral diversity from bulk interaction metrics, while the separation theorems illustrate limits of reduced-form modeling in the presence of structural heterogeneity, analogous to adaptive adversaries behaving differently under interventions. However, practical application would require empirical validation and adaptation to the bot-defense domain's characteristics, especially regarding policy learning and heterogeneous behavior modeling. Overall, the paper offers a mathematically rigorous foundation for thinking about reliability and identifiability in simulations or models relying on population heterogeneity and interacting agents.
Cite
@article{arxiv2607_04627,
title={ Reliability and Identifiability in Persona-Trained Monte Carlo: Variance Decomposition, Stability Bounds, and the Identifiability of Heterogeneous News Reaction },
author={ Salavat Ishbulatov },
journal={arXiv preprint arXiv:2607.04627},
year={ 2026 },
url={https://arxiv.org/abs/2607.04627}
}