Estimating Permanent Price Impact via Machine Learning

In this paper, we show that vector auto-regression (VAR) models, which are commonly used to estimate permanent price impact, are misspecified and can produce conflicting and incorrect inferences when the price impact function is nonlinear. We propose an alternative method to estimate permanent price impact by modifying a reinforcement learning (RL) framework. Our approach assumes the data is stationary and Markov, but is otherwise unrestrictive. We obtain empirical estimates for our model using an iterative learning rule and demonstrate that our model captures nonlinearities and makes correct inferences.