Generating Time Series by Matching Random Convolutional Features

Generating realistic financial time series is challenging as training data is typically limited to a single historical path, making adversarial training prone to discriminator overfitting. Instead, recent work replaces the trained discriminator with a fixed feature map, and trains to generate paths whose features match those of real time series. As features, most prior work chooses path statistics based on signatures from rough path theory. Motivated by strong empirical results of random convolutional features on time series classification, we train generative models with SOCK (**Softmax-Variance Of Competing Kernels) a novel random convolutional feature map that is fully differentiable, highly scalable, and simple to implement. SOCK consistently outperforms other convolutional or signature-based feature maps on hypothesis testing benchmarks and on training generative models.