Dynamic Discrete Choice models with High-Dimensional Proxy Variables [Draft]
Abstract. This paper proposes a novel framework for addressing unobserved heterogeneity in dynamic discrete choice models (DDCMs) using proxy variables. I establish a new nonparametric identification result for models with continuous, time-varying latent state variables by exploiting information contained in proxy variables. To accommodate high-dimensional proxy variables, I develop a flexible estimation procedure that integrates neural network sieves with the EM algorithm. I then establish the consistency and asymptotic normality of the proposed estimator. The framework allows for rich sources of proxy information, including high-dimensional unstructured data, thereby providing a theoretically grounded approach to incorporating such information into dynamic structural models.
SPSA and HARP Algorithms for Constrained Optimization
Abstract. This paper applies the Stochastic Approximation (SA) algorithm to solve black-box constrained optimization problems. It provides convergence results and asymptotic theory for SA algorithm estimates under two scenarios: (i) when the true parameter lies on the boundary of the constraint set and (ii) when it is an interior point of the constraint set. Using the monopolist’s dynamic pricing problem as an example, this paper demonstrates that the SA algorithm’s price estimates converge to the optimal price and asymptotically follow a truncated normal distribution when the optimal price lies on the boundary of the price constraints. Simulation results confirm the convergence and asymptotic theory of the SA algorithm estimates.
Supra-competitive pricing using the SPSA algorithm
Abstract. This paper demonstrates that when two sellers engage in price competition using the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm, collusion can be facilitated through the strategic tuning of hyperparameters. The paper presents scenarios in which prices converge to collusive, supra-competitive, and competitive levels, respectively. While the algorithm does not autonomously learn to collude in a reinforcement-learning sense, these results illustrate how market participants can readily leverage specific algorithmic frameworks to identify and maintain collusive price points.