Analysis and optimization of the time series data with deep artificial neural networks: Financial estimation algorithms

Time series data is analysed with different methods in terms of statistical / mathematical analysis, signal processing, feature extraction, pattern recognition, machine learning and deep learning methods. By analysing the time series data, future values of the data can be estimated. In the analysis of financial time series data, as financial technical analysis indicators are used, machine learning and deep learning algorithms are also used. However, models that integrate technical analysis data with deep neural networks are rarely seen in literature. With the proposed thesis, as a contribution to literature, technical analysis data and deep convolutional neural network are combined. The difference between the proposed models and the existing methods can be explained as follows: Technical analysis data is applied on the prices to create feature vectors and matrices (two-dimensional images) and the financial time series data is converted into a classification problem. In this thesis, our aim is to develop methods that can make financial forecasts in the medium and long term, making stable decisions that can provide maximum profit. In line with these objectives; financial technical analysis indicators, genetic algorithm, deep multilayer perceptron, deep convolutional neural network were used to develop novel algorithms and methods. Four different models were proposed in the thesis. The proposed algorithms were run in a real financial evaluation scenario and results were compared with Buy&Hold strategies, RSI and SMA models, LSTM and MLP regression methods. It has been observed that the obtained results provide better performance when compared to the widely used Buy&Hold models and machine learning models in the literature. Proposed models can be adapted to different time series to be utilized in various use cases.