Good generative models for short time series data exist and have been applied for both data augmentation and privacy protection purposes in the past. A common theme for existing generative models is that they all use a recurrent neural network (RNN) architecture, which makes the models limited regarding the length of the sequences. In real world problems, we might have to deal with data containing longer sequences, and it is such data we in this thesis attempt to synthesize. By combining the recently successful TimeGAN framework with a temporal convolutional network component architecture, we generate synthetic sequential data for two toy data sets: sequential MNIST and multivariate sine waves. The results strongly indicate, although relying solely on a visual inspection, that the model manage to capture long temporal dynamics over time and also relations between different features for the multivariate sine waves data set. In order to make our model applicable for real world data sets, we suggest two improvements. Firstly, the validation of the generated data should not only rely on visual inspection, but also ensure that the synthetic data has the same statistical distribution. Secondly, depending on the task, model refinements such that the synthetic samples look even more realistic should be made.

This study aims to enhance the accuracy of a demand forecasting model, XGBoost, by incorporating synthetic multivariate restaurant time series data during the training process. The research addresses the limited availability of training data by generating synthetic data using TimeGAN, a generative adversarial deep neural network tailored for time series data. A one-year daily time series dataset, comprising numerical and categorical features based on a real restaurant's sales history, supplemented by relevant external data, serves as the original data. TimeGAN learns from this dataset to create synthetic data that closely resembles the original data in terms of temporal and distributional dynamics. Statistical and visual analyses demonstrate a strong similarity between the synthetic and original data. To evaluate the usefulness of the synthetic data, an experiment is conducted where varying lengths of synthetic data are iteratively combined with the one-year real dataset. Each iteration involves retraining the XGBoost model and assessing its accuracy for a one-week forecast using the Root Mean Square Error (RMSE). The results indicate that incorporating 6 years of synthetic data improves the model's performance by 65%. The hyperparameter configurations suggest that deeper tree structures benefit the XGBoost model when synthetic data is added. Furthermore, the model exhibits improved feature selection with an increased amount of training data. This study demonstrates that incorporating synthetic data closely resembling the original data can effectively enhance the accuracy of predictive models, particularly when training data is limited.