Computing Nash equilibria for product design based on hierarchical Bayesian mixed logit models

Abstract: Despite a substantial body of theoretical and empirical research in the fields of conjoint and discrete choice analysis as well as product line optimization, relatively few papers focused on the simulation of subsequent competitive dynamics employing non-cooperative game theory. Only a fraction of the existing frameworks explored competition on both product price and design, none of which used fully Bayesian choice models for simulation. Most crucially, no one has yet assessed the choice models' ability to uncover the true equilibria, let alone under different types of choice behavior. Our analysis of thousands of Nash equilibria, derived in full and numerically exact on the basis of real prices and costs, provides evidence that the capability of state-of-the-art mixed logit models to reveal the true Nash equilibria seems to be primarily contingent upon the type of choice behavior (probabilistic versus deterministic), regardless of the number of competing firms, offered products and features in the market, as well as the degree of preference heterogeneity and disturbance. Generally, the highest equilibrium recovery is achieved when applying a deterministic choice rule to estimated preferences given deterministic choice behavior in reality. It is especially in the latter setting that incorporating Bayesian (hyper)parameter uncertainty further enhances the detection rate compared to posterior means. Additionally, we investigate the influence of the above factors on other equilibrium characteristics such as product (line) differentiation.

• The paper introduces a novel hierarchical Bayesian mixed logit framework to compute Nash equilibria in competitive product design.
• It shows that deterministic (first choice) rules enhance equilibrium recovery, yielding lower prices and greater configuration stability compared to probabilistic rules.
• Comprehensive Monte Carlo simulations underscore that correct behavioral rule specification is crucial for reliable market equilibrium predictions and strategic product design.

Computing Nash Equilibria for Product Design via Hierarchical Bayesian Mixed Logit Models

Introduction

This paper addresses a notable lack of research at the intersection of competitive product design optimization and advanced discrete choice modeling. Specifically, it formulates and computationally investigates the capability of hierarchical Bayesian mixed logit models (HB MXL) in predicting competitive equilibria—Nash equilibria—over both price and product design in markets with differentiated products. The work is characterized by its methodical simulation and its systematic comparison of estimated equilibria to "ground truth" solutions derived from simulated respondent-level preferences.

Literature and Gap Analysis

Prior research integrating conjoint analysis and game-theoretic simulations (notably Choi & DeSarbo 1993, Green & Krieger 1997, Allenby et al. 2014) has progressed from segment- or aggregate-level modeling of consumer heterogeneity to full-fledged individual-level estimation, with some recent attention toward Bayesian methodologies. However, the extant literature neither assessed the efficacy of these models for true equilibrium recovery nor rigorously analyzed the impact of choice rule specification (deterministic vs. probabilistic), nor compared the effect of utilizing full Bayesian posterior draws as opposed to means in equilibrium computation.

This paper's experimental design explicitly fills these gaps, offering for the first time a comprehensive Monte Carlo analysis of how choice model estimation and behavioral rule selection affect the accuracy and properties of Nash equilibria recovered in simulated competitive product line decisions.

Methodological Framework

The methodology is notable for its technical rigor:

• Synthetic Data Generation: Individual-level utilities are sampled from empirically validated preference distributions featuring realistic heterogeneity and monotonicity, limiting potential biases from unrealistic parameter settings.
• Experimental Design: Near-D-optimal choice experiments are generated using Fedorov’s algorithm, maximizing estimation efficiency subject to practical constraints.
• Choice Simulation: Noise is added to deterministic utilities via a controlled Gumbel error process, calibrated using a novel metric (MRGE) to yield interpretable noise-to-signal proportions. Out-of-sample validation is performed via both individual-level hit rates and aggregate share-of-choice RMSE.
• Choice Model Estimation: Preferences and their Bayesian hyperparameters are estimated via MCMC using a hierarchical MNL model. Convergence is carefully assessed (Gelman-Rubin diagnostics), with post-hoc monotonicity enforcement to align with the predefined constraints on price preferences.
• Game-Theoretic Simulation: The market competition is modeled as discrete, closed-loop, multi-stage Nash games, with all product configurations and best-responses exhaustively computed via pre-optimized and cached scenario matrices (leveraging both CPU and paralleled C++/R implementations). Equilibria are identified as fixed-points where no unilateral deviation is profitable.
• Experimental Factors: The simulation systematically varies features, number of firms, product line size, preference heterogeneity, error magnitude, choice rule (first vs. logit rule), and whether full Bayesian draws or only means are used, yielding a broad landscape of market scenarios and estimation paradigms.

Results

Model Recovery and Predictive Accuracy

Under increased heterogeneity and noise, parameter recovery and predictive accuracy metrics degrade as anticipated, consistent across all estimation regimes. Homogeneity, low noise, and lower dimensionality (number of features) consistently enhance estimation and prediction fidelity.

Equilibrium Existence and Structure

• Equilibrium Recovery: The majority of simulated games yield unique Nash equilibria, although the probability of non-convergence (cycles or absence of equilibrium) increases under first choice rules and single-product lines, especially as market dimensionality grows.
• Rounds to Convergence & Cycles: Most equilibria are found within two or three best-response rounds, confirming the practical tractability of the implemented computation even in moderate complexity scenarios.

Key Findings on True Equilibrium Recovery

• Deterministic (First) Choice Rule: When consumer choices in reality are deterministic, the first choice rule applied to posterior draws optimizes equilibrium recovery. Incorporating full Bayesian preference (hyper)parameter uncertainty outperforms using posterior means only.
• Probabilistic (Multinomial Logit) Choice Rule: The logit rule, when matched to true choice behavior, is more robust to errors but less stable in terms of prices and design configuration when recovering true equilibria. Means and draws yield similar, but generally inferior, equilibrium recovery compared to deterministic choice for the equivalent behavioral setting.
• Price Outcomes: Deterministic rules consistently predict lower equilibrium prices and higher stability, whereas the logit rule leads to higher prices and more volatile product designs.
• Product Differentiation: Increased competitor count raises the frequency of differentiated equilibria. Share of differentiated equilibria is similar between estimated and true parameter sets, supporting the assertion that aggregate (structural) equilibrium properties are sensibly estimated even when micro-level recovery is incomplete.
• Contribution Margins: The logit rule's higher equilibrium prices are reflected in higher contribution margins, with broad alignment between true and estimated parameter sets only under the first choice rule.

Implications

For Modelers:

• The equilibrium correspondence between true and estimated models depends critically on a correct choice behavior assumption; misspecification can lead to systematic errors in predicted market prices, design configurations, and profit projections.
• When simulating market competition for managerial decision-support, failure to account for choice rule appropriateness can severely bias strategic recommendations, especially in settings demanding fine-grained price optimization.

For Practical Computing:

• Exhaustive Nash computation remains tractable only for modest market sizes; the scalability bottleneck is not in the optimization of the best-responses per se, but rather in the combinatorics of the initial state space.
• Full Bayesian simulation (using posterior draws, not just means) is computationally demanding but delivers measurable gains in equilibrium recovery rates, especially in deterministic settings.

Theoretical Consequences:

• The structure of equilibria (especially product differentiation and price stability) is more a function of choice rule correctness than of estimation noise or parameter uncertainty.
• Hierarchical Bayesian mixed logit models, given sufficient computational resources and correct behavioral assumptions, have strong capacity to recover not only shares and preferences but also market equilibria.

Limitations and Future Directions

The study is limited to symmetric competitors and does not address scenarios with major cost structure differences, advanced supply-side constraints, Stackelberg games, or integrated no-choice options. The paper highlights that future research should focus on:

• Asymmetric market structures,
• Greater prominence of non-price attribute costs,
• Incorporation of segment-level heterogeneity/flexible mixture models,
• Stackelberg and multi-stage equilibrium concepts,
• Algorithmic breakthroughs for pruning or approximating the vast initial state space without loss of equilibrium information.

Conclusion

This work demonstrates that hierarchical Bayesian mixed logit models are capable of supporting equilibrium-based competitive product design simulation, provided that the behavioral choice rule imposed matches the underlying decision process. Deterministic (first choice) settings benefit substantially from incorporating full posterior uncertainty, with strong numerical recovery and stability of both prices and configurations. For probabilistic settings (logit rule), equilibrium recovery is robust but exhibits greater variance in predicted strategies. Overall, the correct alignment of behavioral rules with market reality is the dominant factor in reliable equilibrium prediction, with significant implications for both theory and practical product management using advanced discrete choice models (2512.22864).