Stock generation estimation using financial news with new generation deep contextualized word display and deep learning models

Objective Stocks are an important investment type affected by the economic crisis. Therefore, it is important for investors, analysts and researchers to predict the direction of the shares. In particular, it is an important source in determining the direction of investments to be made to investors. Users who invest in shares and share their comments on their investments, analysts analyzing shares, and platforms where financial news are published form a platform that provides information sharing to all users. The aim of this study is to propose to the people using the new generation of word embedding models as well as traditional deep learning and word embedding models to predict the direction of the largest volume of shares in BIST100 and to provide investors with an important resource in determining the direction of their investments. To the best of our knowledge, it is the first study to analyze the largest volume of shares in BIST100 using traditional Turkish embedding and deep learning models as well as new generation of word embedding models over completely Turkish texts. Materials and Methods Individual and corporate user reviews, announcements on news sites and financial technical analysis, which is a valuable resource for investors, have been collected as the Turkish text source. Individual and corporate user comments were collected from the accounts by searching on the Twitter pages ("AKBNK", "ALBRK", "GARAN", "HALKB", "ISCTR", "SKBNK", "TSKB", "VAKBN", "YKBNK"). . It was collected by using Selenium Crawler, which we wrote in Python programming language, in order to collect user comments on the social media platform Twitter. With our own web browser in C #, financial news from the Public Disclosure Platform (KAP) and user comments from the Mynet Finans website are collected as various Turkish text sources. Financial analyzes conducted by analysts belonging to Big Para were collected daily. The data in Twitter, KAP and Mynet Finans were collected between 01.09.2018 and 01.09.2019. Since the historical data of Big Para could not be collected, it was collected daily between 28.08.2019 and 15.11.2019. In this study, Word2Vec, GloVe and FastText are used as traditional word embedding models to enrich user interpretations, financial analysis and news in terms of semantic, contextual and syntax. Conventional neural networks (CNNs), Recurrent Neural Networks (RNNs) and Long Short Term Memory Networks (LSTMs) have been implemented with traditional deep learning algorithms for the classification task. In addition, the new generation of word embedding models from the Transformers Bidirectional Encoder Display (BERT), Language Models Placement (ELMo) and Universal Language Model Fine Tuning (ULMFiT) were used. Results In this study, extensive experiments have been conducted to predict the direction of large volume stock market shares in BIST100 by using traditional word embedding models, deep learning algorithms and next generation word embedding models. All stated accuracy is an evaluation criterion used in experiments to demonstrate the classification performance of each model and the contribution of our work. With the application of pre-treatment methods, it is aimed to improve the classification performance of the proposed model. The combination of ELMo, which is a new generation word embedding model for classifying Turkish texts containing user comments, with preprocessing methods, is an advantageous choice for determining the sensitivity of the users in guiding their shares and achieving the best classification success. 97.70% and 91.55% with the accuracy value was revealed. However, traditional deep learning algorithms produced better results than the new generation word embedding models in Turkish textual data sets such as news and analysis. Conclusions This study demonstrates the effectiveness of using traditional word embedding models, deep learning algorithms, and new generation word embedding models on texts collected from various data sources to predict the direction of stock market shares and makes a valuable contribution to investors in the process of investing.