WP2
Our findings show that while there are tools available to handle sentiment analysis extracted from social networks,the mostly show positive/negative sentiment or positive/neutral/negative. Apart from it being difficult to audit how these ‘labels’ were given, such binary granularity is of no use for quantitative modelling of financial data. What makes SSIX unique is that it lays foundations to generate ‘technical sentiment data’ parameters or ‘X-scores’ which are within a continuous numeric interval (-1;1) having the potential to be adopted by the quantitative investment industry and reach the same level of utility as commonly used parameters such as P/E ratio, RSI or MACD.
WP3
Performance tests highlighted stability issues due to the high volumes of parallel data when listening to the most discussed financial markets, however, this bottleneck was overcome with hardware resources scalability. Our experiments shifted to cloud technologies provided by the Google Cloud Platform reduced the time for data storage and extraction and helped the scalability of the parallel computing processes. A stratified sampling technique was developed to extract content from large historical data sets. This technique was adopted for creating the data sample used in the production of the platforms custom classifiers.
WP4
By the first year, we concluded that the available multilingual domain specific and sentiment lexica may not provide the expected features for the opinion mining needs for this project. In parallel, several Big Data analysis infrastructures were analysed for their suitability used as the foundation for the pipeline architecture. Year 2 and 3 saw development on SSIX custom sentiment classifiers for financial microblogs, the Brexit referendum and the 2017 German elections. Benchmarking tests for the financial microblogs classifier show promising results against current SOTA services. Efforts have gone into a custom machine translation service and an aspect-based sentiment analysis (ABSA) classifier.
WP5
Foresight was given to the need for scalability and efficiency. The major system components were designed to operate independently of each other, so they can be distributed or centralised depending on the deployment scenario and load on the system. Areas of potential innovation include testing of new classification models, building a system for statistical calculations and NLP classification, using massively parallel computing and researching new visualisations to aid end users in the decision-making process.
