Please use this identifier to cite or link to this item:
Title: Financial news sentiment analysis
Authors: Parnis, Mario
Keywords: Social media
Machine learning
Kernel functions
Issue Date: 2018
Citation: Parnis, M. (2018). Financial news sentiment analysis (Bachelor's dissertation).
Abstract: The rise in use of social media in recent years, has led to a community of traders to emerge on microblogging platforms where they share their thoughts, opinions and in-sights publicly. Building good techniques to analyse such data and extract meaningful sentiment from it can lead to a market advantage. This project deals with analysing different data-driven techniques using machine learning to bring-to-light what techniques work and what needs to be improved upon. As part of the research three machine learning models are created: an N-Gram Bag-of-Words, POS pattern Bag-of-Words and Word Embeddings based models. These are paired with a Support Vector Regressor and different variants of Na•ve Bayes. As an aid to future researchers all models created in project have also been packaged and a simple interface has been provided to allow them to be used in other projects. The results produced show that domain-specific systems are a must for microblog financial news sentiment analysis. Moreover, the top performing model which makes use of N-Gram Bag-of-Words and Support Vector Regression, produces a score of 0.7227 Weighted Cosine Similarity (WCS). This result is very close to state-of-the-art systems which utilise more complex features and techniques to get a small boost in performance. This signifies that current systems are not able to achieve results much above baseline and that further research will be required to bring this area forward. In this regard, the objective of this project to analyse current techniques has been achieved.
Appears in Collections:Dissertations - FacICT - 2018
Dissertations - FacICTAI - 2018

Files in This Item:
File Description SizeFormat 
  Restricted Access
809.21 kBAdobe PDFView/Open Request a copy

Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.