Analysis of graph time series with deep learning

Time series are used to understand the temporal variation of an object. Time series are encountered in various fields such as finance, energy and traffic. Moreover, many problems such as anomaly detection, behavior recognition can be modeled as time series problems. Therefore, the analysis of time series which are frequently encountered in real life is of great importance and time series problems are widely researched and tried to be solved. Graphs are used to analyze the relationships between objects. Some difficult problems can be better understood when the problem data is analyzed as graphs. Hence, graph problems have an important place in the literature. Time series and graphs provide an opportunity to understand different aspects of the problem. Therefore, there are studies in the literature that make better modeling by combining time series and graphs. This method is applied in some studies in the field of finance. In this thesis, graph series analysis was performed with deep learning methods for financial forecasting problems. For this purpose, DOW 30 stock market is represented as a graph. Then, graphs at different timestamps were ordered and graph series was formed. The amount of change in the shares was predicted by training artificial neural networks with obtained series. According to the predicted amount of change, the principal was invested by applying a daily buy/sell strategy. As a result of this, the annual return was calculated as a percentage. As a result of the study, higher average annual returns were obtained compared to deep learning models using only time series and it was concluded that using graph series significantly improved financial forecasting, in other words, inferences that could not be captured with time series could be made. In the thesis study, graph series was created with multiple methods. A similar average annual return was obtained with deep learning models trained with different graphs. Thus, it has been shown that the method of obtaining a series of graphs is a robust and stable method. In addition, an ensemble model that predicts the stock that increases the most from the outputs of the trained deep learning models is trained. An average annual return of 26.68 % was achieved with the final model. It has been shown this method provides more profit than various greedy algorithms in the literature as well as the basic methods in the literature. As a result, the developed ensemble model is proposed as a method that can be used for daily buy/sell strategy in real life.