Analysis of financial data using natural language processing and deep learning methods

In recent years, with the rapid increase in digital data production at worldwide, its usage for decision-making has significantly grown. This trend has also made its mark in the finance. However, traditional statistical methods are no longer effective due to the uncontrolled expansion and complexity of raw data. Therefore, it is crucial to employ modern machine learning methods for cleansing and analyzing financial data. This thesis presents innovative approaches that generate new datasets from annual and periodic financial reports shared by companies and analyze these generated datasets using machine learning techniques. Within the scope of the study, 10K annual reports were collected, transformed into analyzable datasets, and subjected to classification processes using machine learning methods. The successful results obtained (92% accuracy) contribute significantly to the transformation of 10K reports into a dataset. The application of natural language processing techniques to complex and error-prone 10K reports was also performed within the thesis, presenting unique approaches. A hybrid method that enables the analysis of 10Q quarter reports was produced by the studies carried out within the scope of the thesis. With the method, companies' quarterly reports with textual content were transformed into efficient data sets using Doc2Vec and K Means Clustering algorithms. The 'PriceRank' metric, which shows the price power of companies in the next financial quarter and takes values between one and ten, was divided into three groups as low, medium, and high and added to the data set as class values. Later, successful results (84% accuracy) were produced by the classification process with the deep learning method, Convolutional Neural Network. Finally, a hybrid method consisting of 1D Convolutional Neural Network and Long Short-Term Memory algorithms was applied to the data set produced with Doc2Vec and K Means algorithms to better reveal both short- and long-term connections on the data and to improve the feature extraction step. The successful results obtained (88% accuracy) showed that this new hybrid method can produce better results than traditional deep learning methods in data analysis.