Project Summary:

This final 6 week project for the LSE course gave myself and six students the opportunity to work together in a collaborative project aimed at providing strategic insights to The Bank of England. It served as a quasi internship as there was frequent contact and presentations to The Bank itself. The focus was on speech data sentiment scores and any potential correlation to economic indicators. This project had a significant scope so, unlike previous assignments, my personal involvement was being a dedicated co-leader of the team and leading the NLP analysis specifically.

Personal Analytical Approach Highlights:

• Co-leadership: I rose to the challenge of co-leader to assist the original team leader which intensified my responsibilities beyond NLP responsibilities. This included 1:1 management of a struggling team member, initiation and upkeep of all team work/communication platforms and ensuring adherence to best practices and deadlines.

• High level overview: I created an solid, actionable and well researched high level plan to address business questions from a nebulous initial brief in spite of not having previous economic experience.

• NLP: I researched and implemented four different NLP models to explore potential sentiment scores; TextBlob, Vader, Loughran McDonald and FinBERT. This includes the processing of both speech data and FSR, MPC reports across the time period. Descriptive statistics, visualisations and proactive discussion were used in evaluation of the models to refine the final choice down to FinBERT.

  • A time series of the average sentiment scores by year over time across all models

• Presentation: I effectively communicated our findings to stakeholders at The Bank of England, demonstrating my ability to distil complex analyses into actionable insights.

Personal Key Insights and Recommendations:

• I identified, tested and evaluated two general and two financial NLP models in order to provide a useful collection of models. In particular the distribution of sentiment scores helped to refine models down to a clear choice.

  • An analysis of the sentiment score distribution across all models

• I explored word frequencies for most positive/negative speeches in order to give the Bank a basic starting point to understand the linguistic patterns that contribute to a negative polarity score and to strategise their communication accordingly. Including words like edge Global, Credit, Crisis, Demand and Year exclusively appearing in top negative speeches and not positive ones.

  • A word cloud of most frequent negative words using FinBert as the model

• I observed a lack of significant correlation between sentiment scores and economic indicators by exploring time series averages. This allowed the team to identify significant periods such as The Global Financial Crisis, Brexit or the invasion of Ukraine as key drivers for change in both economic indicators and sentiment scores.

Professional Development and Impact:

This project showcased my interpersonal communication skills through a leadership position. I was also able to take on a variety of administration roles whilst responsible for my own entire section of the project, reinforcing my agile work approach. Finally, a genuine passion for NLP was cemented in the exploration of the business questions, a passion which I would love to be able to explore further.