LEE- Endüstri Mühendisliği-Doktora

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/19268

Gözat

Konu "büyük veri" ile LEE- Endüstri Mühendisliği-Doktora'a göz atma
  • Öge
    Sentiment analysis model proposal with deep learning techniques on big data: Portfolio selection with the help of industry indicators
    (Graduate School, 2023-09-12) Sivri, Mahmut Sami ; Üstündağ, Alp ; 507162108 ; Industrial Engineering
    The prediction of stock movements is a highly complex problem due to the influence of numerous variables on performance. However, even a slight improvement in prediction accuracy can lead to significant impacts on the rate of return. Also, it is very difficult to determine the right variables, methods and parameters that will be used to predict these changes since stock markets are very complex and dynamic. Processing unstructured data has become widespread in the era of big data besides handling relational or structural data. In recent years, machine learning and sentiment analysis have been visited by many researchers to predict the stock market where integration such types of unstructural data, such as news and financial data, has direct impact. While only technical indicators and macroeconomic data were utilized in financial prediction problems in the past, nowadays researches have shown that also using news, comments and reports as data sources gives better outcomes. It has been observed that these non-numerical data make a significant difference to the performance in the stock market prediction. Efficient financial management is a key challenge facing businesses today. One critical aspect of financial management that businesses should prioritize is the creation and maintenance of an investment portfolio. This is especially important in today's globalized society, where intense competition and rapid economic changes at both national and international levels create a complex operating environment. In this thesis, a stock prediction framework is developed with feature selection, prediction and evaluation methods by using a variable pool consisting of different variable groups. Contrary to the existing studies, the change between the opening and closing prices of the next day, which is more suitable for real life, is predicted. Also, daily sliding window cross validation methodology is included in the study to reflect the real-life scenario. The framework consists of ten main and four expanded variable groups ranging from financial to operational indicators. Experimental results showed that competitive performance in terms of accuracy and rate of return were achieved. Detailed presentations of techniques for sentiment analysis and text analytics are included in the next section. Fundamental methods of text analytics such as preprocessing of texts, text-based feature engineering techniques, text classification and topic models are presented. At the end of the chapter, sentiment analysis models are explained with examples. Approaches and results used in different sentiment analysis models are compared. Also, with state of art sentiment analysis models, the sentiment labels of the sector news were predicted and the results from this framework were combined with a fuzzy soft-set method in this study. To determine the fuzzy membership function, the normalized weekly cumulative rate of returns of prediction models were utilized as the criteria. In addition to proposing a new approach, our research varies from previous studies in terms of data coverage and the models that we used in both sentiment analysis and prediction phase. Performance of the proposed methodology was evaluated with the cases where news and financial data are used separately and together Finally, comparative analyzes with portfolio selection and optimization methods are presented. Portfolios created with different models for stocks will be compared with the performance metrics whose definitions are given.