Hands-On Machine Learning for Algorithmic Trading

Synthetic time series with recurrent cGANs

Recurrent (conditional) GANs are two model architectures that aim to synthesize realistic real-valued multivariate time series. The authors target applications in the medical domain but the approach could be highly valuable to overcome the limitations of historical market data.

RGANs rely on RNNs (see the last chapter) for the generator and the discriminator. RCGANs further add auxiliary information in the spirit of cGANs (see the previous section).

The authors of cGAN succeeded in generating visually and quantitatively compelling realistic samples. Furthermore, they evaluated the quality of the synthetic data, including synthetic labels, by using it to train a model with only minor degradation ...