deep learning

We propose a method using temporal convolutional networks for model parameter estimation by learning the mapping from sample data to its generating parameters. This map is then used for defining moment conditions in simulation-based inference. Our approach outperforms maximum likelihood estimators for small and moderate samples and effectively estimates a jump-diffusion financial model.

We conducted a Monte Carlo simulation study on the application of deep reinforcement learning methods for portfolio management with predictable returns and costly transactions. Our findings reveal varying algorithm performance, emphasizing the need to restrict asset samples for sufficient DRL performance. We also contributed to the Julia programming language ecosystem by providing codes for work replication and extension.