Time series analysis and forecasting: Deep learning approach

In daily life, quite a lot of problems include time series data. Making good time series data analysis or future value estimations is very important to solve these problems. Various statistical analysis, mathematical analysis, signal processing, machine learning and deep learning methods that are subfields of machine learning are used to analyze time series data and to make a future forecast. In particular, deep learning methods, which have become increasingly popular in recent years, have been more successful than traditional methods in solving some complex time series problems and their usage has spread rapidly. However, in the literature, there is not a lot of research which deals with the analysis and estimations, how to approach the end-to-end solutions, which models should be used, how to use the selected model and how the dataset should be prepared. In the proposed thesis, various time series problems are examined, and approaches are explained. Stateful LSTM and the stateless LSTM were compared, analyzed on simple problems, internal structures of them were examined. After that, electricity production and consumption amount of a household were estimated, and the battery of household were optimized. When battery optimizations were made, the optimum result was converged at a rate of 99% and a high rate of profit was achieved by the use of 3-rate electricity tariffs. In addition, the technical analysis features calculated over the financial data and used for the estimation of the future value of the financial data with the help of CNN, another deep learning model. In this estimation, the conversion of technical indicators which will be the inputs of CNN model as 2D images is an uninvestigated issue. The dendrogram clustering algorithm used in 2D image rendering provided one-axis correlation, while the other axis was used for autocorrelation of the time series. When the results were examined and transformations were applied, developed method could be trained and gave successful results. Since the transformations applied to finance data are also applicable to time series data in different fields, models which have different architecture can be useful.