This thesis aims to improve the accuracy of a demand forecasting gradient boosting machine learning model, XGBoost, by generating synthetic restaurant time series data and adding it to the training procedure of the model. This was all done under the premise that there existed an initial lack of data to properly train the model. The original data used in the study is a time series dataset comprising numerical and categorical features based on a real restaurant’s sales history, complemented by relevant external data. The synthetic data is created using a generative adversarial deep neural network model, TimeGAN, specifically tailored for time- series data. It learns from the original data and aim to retain its temporal and distributional dynamics in the synthetic data. The results of the study show strong statistical and visual similarity between the synthetic and original data. An evaluative experiment where synthetic data of differing lengths was iteratively concatenated to the one-year real dataset was conducted. For each iteration of added synthetic data, the XGBoost model was retrained and its accuracy on a one-week forecast horizon was evaluated using the Root Mean Square Error (RMSE). Further, the hyperparameters of the model were optimized at each iteration of the experiment using a Tree Parzen Estimator (TPE). The results of the experiment indicated that the addition of up to 6 years of synthetic data improved the model performance by up to 65%. The hyperparameter configurations of each model suggests that, with added synthetic data, the XGBoost model can benefit from deeper tree structures in its decision making. Further, the model seems to become better at feature selection when more data is fed during the training process. Overall, the results suggests that the addition of synthetic data that closely resembles the original data can be useful for training accurate predictive models in cases when training data is limited.