Working Paper: NBER ID: w21124
Authors: Carlos A. Manzanares; Ying Jiang; Patrick Bajari
Abstract: In this paper, we propose a method for finding policy function improvements for a single agent in high-dimensional Markov dynamic optimization problems, focusing in particular on dynamic games. Our approach combines ideas from literatures in Machine Learning and the econometric analysis of games to derive a one-step improvement policy over any given benchmark policy. In order to reduce the dimensionality of the game, our method selects a parsimonious subset of state variables in a data-driven manner using a Machine Learning estimator. This one-step improvement policy can in turn be improved upon until a suitable stopping rule is met as in the classical policy function iteration approach. We illustrate our algorithm in a high-dimensional entry game similar to that studied by Holmes (2011) and show that it results in a nearly 300 percent improvement in expected profits as compared with a benchmark policy.
Keywords: No keywords provided
JEL Codes: C44; C55; C57; C73; L1
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Component-wise gradient boosting estimator (C51) | Reduction in dimensionality of the problem (C29) |
| Avoiding equilibrium restrictions (D59) | More nuanced understanding of strategic interactions (C72) |
| Algorithm generates improvement in expected profits (C69) | Improvement in expected profits (D25) |
| Using past gameplay data (Y10) | More effective decision-making (D91) |