Capstone Project

Iain MacCormick

June 9th, 2020

I. Definition

Project Overview

With a wealth of information and opinion available in real time through news, social media and investor publications there is a vast source of financial textual data, which if understood correctly can be used by investors and traders to better understand companies and markets.

This project will attempt to classify the the sentiment of a Financial statements from the point of view of a retail investors. Since sentiment of financial news articles and headlines, as well as other financial texts, can be an indicator of company stock returns and volatility it is important for retail investors to have access to an overview of sentiment from the financial news towards a company.

Sentiment analysis is an algorithmic approach used to understand the emotional context of a body of text and is used in a wide range of domains including but not limited to: understanding polarity of financial texts, interpreting product reviews and identifying racism in social media posts.

The dataset that will be used for experimental analysis is known as the "Financial Phrase Bank" dataset containing sentences of positive, negative and neutral sentiments. In some of the approaches covered, domain specific dictionaries will be used.

Problem Statement

The aim of this project is to give retail investors access to statistics describing the sentiment of financial news headlines for companies of their interest. To achieve this functionality a Sentiment Classifier will need to be trained, tested and deployed and a web application will need to be built to enable the user to retrieve predictions.

Modelling Steps:

1. Download dataset and conduct exploratory analysis.
2. Conduct a series of experiments with various feature engineering steps and classification algorithms, discussed in algorithms and techniques section.
3. Select the final model based upon performance against our selection criteria discussed in the metrics section.

Web Application:

The selected model will be deployed and integrated within web application to give retail investors access to sentiment analysis for news relating to companies of interest. The purpose of this document is to cover the experimental analysis only however the final application can be accessed here: https://fin-sent.herokuapp.com/

Metrics

We aim to minimise the number of true positives and true negatives so accuracy will be used to evaluate the benchmark and solution model. Precision, Recall and F-Score will be used alongside accuracy to help understand the performance of the model during development. As the problem is a multi-class classification with class imbalance the micro averaging method will be used for these metrics.