Endogeneity Paradox in Markets and Models

• Endogeneity Paradox is a phenomenon where feedback mechanisms and variable interdependence induce systematic biases in models predicting prices, rewards, or causal effects.
• Researchers address the paradox using methods such as local martingale approximations, comparative OLS differences, and instrumental variable corrections in reinforcement learning.
• Implications include heavy-tailed return distributions, biased change score analyses, and reinforcement biases that underline the need for rigorous diagnostics and robust corrective strategies.

The Endogeneity Paradox encompasses a set of phenomena arising when classical statistical, econometric, or financial modeling predicts reward, price, or causal effects in the presence of endogenous relationships, yet observed dynamics or inferences confound or violate expected regularities. It is manifested in at least three distinct research literatures: derivatives pricing in financial economics, causal inference in the analysis of change scores, and dynamic decision-making or sequential learning. Across domains, the paradox is rooted in the tension between conventional identification strategies and endogenous features—market participant behavior, feedback, or variable interdependence—that generate systematic bias or nonstandard dynamics, even in the absence of exogenous shocks or classical sources of statistical confounding.

1. Risk Neutral Valuation Paradox in Derivatives Markets

The "Risk Neutral Valuation Paradox," as formulated by Kardaras (Maccioni, 2011), exemplifies the Endogeneity Paradox in informationally efficient, frictionless, but incomplete derivatives markets. Classical financial theory, guided by the Fundamental Theorem of Asset Pricing, posits that in the absence of arbitrage, and even with market incompleteness, asset prices should evolve as martingales under equivalent risk-neutral measures, reflecting fundamental values derived from expected discounted payoffs.

However, when the market includes both risk-averse (fundamental) traders and risk-neutral (technical) traders, a paradox emerges: price dynamics can systematically and endogenously escape fundamental value zones without any informational shock or irrational trading. Technical traders, by virtue of their ex-ante indifference and subsequent Bayesian updating, submit market orders just beyond current equilibrium, imparting small persistent drifts. This causes market-clearing prices, denoted $\mathrm{Equil}(\mathbf{x}_t)$, to periodically exit the compact interval bounded by fundamental reservation bids and asks, resulting in "bubbles" or "crashes" purely via endogenous mechanisms. These are not permanent; upon technical traders' withdrawal, prices jump discontinuously back to the Pareto-efficient fundamental range.

The model, formalized via local martingale and semimartingale approximations, supports the following:

• Martingale price dynamics under technical trading: $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$, with $B_t$ a $\mathcal P$–local martingale and $\int_0^t H_s ds$ of finite variation.
• Girsanov re-weighting: Ensures $Z_t$ becomes a local martingale under an equivalent risk-neutral measure $\mathcal Q$, preserving martingale dynamics but enabling drift in the physical measure.
• Cyclic price deviations: Cycles between "normal" (fundamental) and "bubble/depression" (technical) pricing occur almost surely in sufficiently long horizons.
• Return distributions: The mixture of continuous martingale evolution with randomly timed jumps yields heavy-tailed return distributions, contrasting with the classical Gaussian random walk.

This risk-neutral valuation paradox demonstrates that the machinery designed to enforce market rationality and informational efficiency, under plausible real-world market frictions (e.g., incompleteness), results in endogenous price instabilities that evade classical arbitrage or information-based corrections (Maccioni, 2011).

2. Comparative Studies and the Endogeneity Paradox in Regression

The statistical version of the Endogeneity Paradox is evident in comparative regression contexts, as formalized in Sun and Lysy (Kashyap, 2022). In standard linear modeling, $y = X\beta + u$, endogeneity is typified by $E[X'u]\neq0$, which biases ordinary least squares (OLS) estimators, rendering $E[\hat{\beta}]\neq\beta$. Instrumental variables (IV) are the classical remedy, but the paradox arises when considering comparative inference:

Despite individual OLS coefficients being biased due to persistent endogeneity, the difference in OLS coefficient estimates across two periods or systems—denoted $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$0—is unbiased for the change in structural parameters $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$1, provided a simple condition holds: the product of the covariance structure between explanatory variables and the covariance between error and explanatory variables must be invariant across samples.

Formally, if

$$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$2

then

$$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$3

Notable implications:

• Alternative to IV: Comparative OLS difference estimators can recover true structural changes without external instruments, as long as endogeneity mechanisms are stable across samples.
• Key assumption: Covariance-equality across samples. Violation of stability in endogeneity (e.g., systematic change in omitted variable bias) precludes unbiased inference.
• Practical diagnostics: Estimation relies on hypothesis tests or robustness checks of estimated covariance terms $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$4 across samples.

This result expands the conceptual toolkit for comparative studies where instrument selection is infeasible or instruments are weak, provided the underlying bias structure remains invariant (Kashyap, 2022).

3. Reinforcement Learning, Feedback, and the Endogeneity Paradox

The sequential decision-making domain, particularly in reinforcement learning (RL), exhibits a dynamic form of endogeneity paradox, as articulated by Li, Luo, and Zhang (Li et al., 2021). Here, the agent’s policy both exploits and generates data, leading to a reinforcement bias—distinct from static endogeneity—when estimation of reward parameters is performed from data generated by the (possibly suboptimal) learned policy.

For an infinite-horizon Markov Decision Process (MDP) $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$5, with observed reward $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$6 and endogeneity $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$7, classical OLS policy evaluation is biased, and repeated feedback updates compound this bias via a fixed-point reinforcement mechanism. The limiting learned parameter is

$$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$8

where $$\mathrm{Equil}(\mathbf{x}_t)\approx Z_t = B_t + \int_0^t H_s ds$$9 captures the magnitude of endogeneity-induced feedback.

Resolution employs IV-based reinforcement learning algorithms, which embed the instrument framework into the stochastic approximation backbone of RL (e.g., Q-learning, Actor–Critic):

• Instrument conditions: $B_t$0 must satisfy relevance $B_t$1 full rank and exogeneity $B_t$2.
• Algorithmic correction: Two-stage updates recover the true projected Bellman equations, eliminating R-bias and restoring parametric rates of convergence and valid statistical inference.
• Theoretical guarantees: Asymptotic normality and steady-state consistency of optimal policy parameter inference.

This application operationalizes the endogeneity paradox as a reinforcement learning pitfall, generalizes it to data-generating processes with feedback, and validates IV-based correction as necessary for inferential validity in policy learning under endogeneity (Li et al., 2021).

4. Change Score Analysis, Lord’s Paradox, and Endogeneity

Lord’s Paradox, as dissected by Tennant et al. (Tennant et al., 2023), epitomizes a statistical endogeneity paradox in the analysis of change scores from observational data. With two outcome measures $B_t$3 (baseline) and $B_t$4 (follow-up), the observed change $B_t$5 contains three components: endogenous (scale) change, exogenous (causal) change from an exposure $B_t$6, and random noise.

Two common analytic strategies are:

• Approach 1 ("change score regression"): Regress $B_t$7 on $B_t$8.
• Approach 2 ("ANCOVA"): Regress $B_t$9 on $\mathcal P$0 and $\mathcal P$1.

Neither approach reliably isolates the causal effect of $\mathcal P$2 (the exogenous component) without strong, usually unattainable, assumptions:

• Change score regression is unbiased only if $\mathcal P$3, which is rarely satisfied due to regression to the mean and selection biases.
• Regression adjusted for $\mathcal P$4 targets the controlled direct effect of $\mathcal P$5 on $\mathcal P$6 at fixed $\mathcal P$7, not the total causal change, and is vulnerable to mediator-outcome confounding, regression dilution, and collider bias.

The resulting paradox is that standard difference score or baseline-adjusted regressions fail to recover the total exogenous causal effect of $\mathcal P$8—the inference is confounded by endogenous dynamics (scale change), unmeasured noise, or complex feedback structures. Valid inference demands multivariate causal methods, such as longitudinal g-formula, inverse probability weighting of marginal structural models, or structural nested models, and typically requires more than two outcome waves to distinguish endogenous from exogenous and random components (Tennant et al., 2023).

5. Mechanisms and Implications Across Domains

The Endogeneity Paradox, encountered in financial markets, causal inference, and reinforcement learning, underscores a common methodological challenge: endogenous mechanisms—be they market participant feedback, structure in error-repressor covariances, or policy-data feedback—systematically invalidate naïve statistical inferences or market predictions.

Key cross-disciplinary implications include:

• Endogenous trading, policy, or change mechanisms can generate heavy-tailed distributions, boom-bust cycles, or unpredictable jumps even in informationally efficient, frictionless, or rational environments (Maccioni, 2011).
• Recovery of unbiased causal, policy, or comparative parameter changes often requires stronger invariance or exogeneity assumptions, monitoring of bias-inducing structures, or the introduction of external sources of variation (instruments).
• The paradox cautions against uncritical adoption of change scores, period-to-period functional inferences, or price–value mapping in models where the underpinnings of endogeneity are not explicitly controlled or tested (Kashyap, 2022, Tennant et al., 2023).
• Statistical remedies (IV estimation, covariance checking, g-methods) must be accompanied by rigorous diagnostic assessment of underlying structural or error dependencies.

A plausible implication is that the Endogeneity Paradox, rather than being an edge-case or theoretical curiosity, is pervasive in modern empirical designs—market microstructure, comparative regressions, longitudinal causal inference, and data-driven sequential decision frameworks alike—necessitating principled methodological vigilance.