Prediction of Time Series Using Generative Adversarial Networks

Abstract

The GAN model consists of an LSTM as the time series generator and an ANN as the discriminator, using the simple moving average and exponentially weighted moving average results as input features for the GAN network, followed by the Fourier transform, ARIMA to create the input features, and finally XGBoost to filter the final prediction data. The GAN network model is generally used for adversarial image generation, and the GAN adversarial network is usually trained as two separate and alternating networks: the recognition network is trained first, then the generation network, then the recognition network, and so on, until a Nash equilibrium is reached. The power of GAN is that it can automatically define the potential loss function. The discriminatory network can automatically learn a good discriminant, which is equivalently understood as learning a good loss function to compare good or bad discriminant results. Although the overall loss function is still artificially defined, the discriminant network potentially learns the loss function hidden in the network, which varies from problem to problem, so that the potential loss function can be learned automatically. Using this particular property of the week to predict time series would be a new approach.