Time series (re)sampling using Generative Adversarial Networks.

Area of Science:

• Statistics
• Machine Learning
• Time Series Analysis

Background:

• Traditional time series analysis often relies on resampling methods.
• Generative Adversarial Networks (GANs) have shown potential in learning complex data distributions.

Purpose of the Study:

• To propose a novel bootstrap procedure for time series data using GANs.
• To evaluate the performance of GAN-based resampling against traditional methods.

Main Methods:

• Utilizing Generative Adversarial Networks (GANs) with temporal convolutional neural networks for generator and discriminator.
• Training GANs on single time series sample paths to generate new samples.
• Comparing GAN-based bootstrap with circular block bootstrapping via simulations.

Main Results:

• GANs can effectively learn the dynamics of stationary time series processes.
• GANs trained on a single sample path can generate additional, convincing samples.
• The proposed GAN bootstrap method can outperform circular block bootstrapping in empirical coverage for AR(1) processes.

Conclusions:

• GANs offer a promising new approach for time series resampling.
• The novel GAN bootstrap procedure demonstrates competitive or superior performance compared to existing methods.
• The method has practical applications, as shown in the Sharpe ratio analysis.