INCOMPLETE TIME SERIES FORECASTING USING GENERATIVE NEURAL NETWORKS

Abstract

Dealing with missing data is a long pervading problem and it becomes more challenging when forecasting time series data because of the complex relationships between data and time, which is why incomplete data can lead to unreliable results. While some general-purpose methods like mean, zero, or median imputation can be employed to alleviate the problem, they might disrupt the inherent structure and the underlying data distributions. Another problem associated with conventional time series forecasting methods whose goal is to predict mean values is that they might sometimes overlook the variance or fluctuations in the input data and eventually lead to faulty predictions. To address these issues, we employ a probabilistic forecasting technique that can accommodate the variations in data and predict a full conditional probability distribution of future values given past data. We introduce a novel generative adversarial network (GAN) architecture with the goal to forecast a probability distribution on time series data and also introduce an auxiliary GAN which learns the temporal pattern of which data is missing, thereby removing the dependency on using general-purpose imputation methods. We create two complex time series datasets to test our architecture and also show a comparison between our architecture’s forecasting capability (with incomplete data) to a state-of-the-art architecture that is trained with complete data. We also demonstrate that our model’s predicted data distribution does not collapse with incomplete data, but instead successfully learns to estimate the true underlying data distribution.