A Smart Data Ecosystem for the Monitoring of Financial Market Irregularities

Evans, Lewis (2022) A Smart Data Ecosystem for the Monitoring of Financial Market Irregularities. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (18MB) | Preview

Abstract

Investments made on the stock market depend on timely and credible information being made available to investors. Such information can be sourced from online news articles, broker agencies, and discussion platforms such as financial discussion boards and Twitter. The monitoring of such discussion is a challenging yet necessary task to support the transparency of the financial market. Although financial discussion boards are typically monitored by administrators who respond to other users reporting posts for misconduct, actively monitoring social media such as Twitter remains a difficult task. Users sharing news about stock-listed companies on Twitter can embed cashtags in their tweets that mimic a company’s stock ticker symbol (e.g. TSCO on the London Stock Exchange refers to Tesco PLC). A cashtag is simply the ticker characters prefixed with a ’$’ symbol, which then becomes a clickable hyperlink – similar to a hashtag. Twitter, however, does not distinguish between companies with identical ticker symbols that belong to different exchanges. TSCO, for example, refers to Tesco PLC on the London Stock Exchange but also refers to the Tractor Supply Company listed on the NASDAQ. This research has referred to such scenarios as a ’cashtag collision’. Investors who wish to capitalise on the fast dissemination that Twitter provides may become susceptible to tweets containing colliding cashtags. Further exacerbating this issue is the presence of tweets referring to cryptocurrencies, which also feature cashtags that could be identical to the cashtags used for stock-listed companies. A system that is capable of identifying stock-specific tweets by resolving such collisions, and assessing the credibility of such messages, would be of great benefit to a financial market monitoring system by filtering out non-significant messages. This project has involved the design and development of a novel, multi-layered, smart data ecosystem to monitor potential irregularities within the financial market. This ecosystem is primarily concerned with the behaviour of participants’ communicative practices on discussion platforms and the activity surrounding company events (e.g. a broker rating being issued for a company). A wide array of data sources – such as tweets, discussion board posts, broker ratings, and share prices – is collected to support this process. A novel data fusion model fuses together these data sources to provide synchronicity to the data and allow easier analysis of the data to be undertaken by combining data sources for a given time window (based on the company the data refers to and the date and time). This data fusion model, located within the data layer of the ecosystem, utilises supervised machine learning classifiers - due to the domain expertise needed to accurately describe the origin of a tweet in a binary way - that are trained on a novel set of features to classify tweets as being related to a London Stock Exchange-listed company or not. Experiments involving the training of such classifiers have achieved accuracy scores of up to 94.9%. The ecosystem also adopts supervised learning to classify tweets concerning their credibility. Credibility classifiers are trained on both general features found in all tweets, and a novel set of features only found within financial stock tweets. The experiments in which these credibility classifiers were trained have yielded AUC scores of up to 94.3. Once the data has been fused, and irrelevant tweets have been identified, unsupervised clustering algorithms are then used within the detection layer of the ecosystem to cluster tweets and posts for a specific time window or event as potentially irregular. The results are then presented to the user within the presentation and decision layer, where the user may wish to perform further analysis or additional clustering.