Four ways to apply NLP in financial services

Natural language processing (NLP) is increasingly used to review unstructured content or spot trends in markets. How is Refinitiv Labs applying NLP in financial services to meet challenges around investment decision-making and risk management?

1. Natural language processing models can be trained to review unstructured content, and spot issues or trends that may impact financial markets.
   
2. Content enrichment and sentiment analysis can help financial institutions make more informed investment decisions, and streamline risk management and compliance, especially in response to COVID-19.
3. NLP in financial services is a key focus for Refinitiv Labs, with work ongoing to quantify sentiment on more than 100 key drivers of equity performance across different content types.

Refinitiv Labs leverages natural language processing (NLP) to optimize data curation, enrich unstructured content, and improve content workflows and data management.

This article looks at some of the benefits of applying NLP in financial services, as well as practical use cases, including Refinitiv Labs projects described to me by Kelvin Rocha, Lead Data Scientist at Refinitiv Labs.

Tackling a firehose of information is a familiar problem in the financial services industry.

Traders and investment managers have numerous sources to comb through, such as research reports, company filings, and transcripts of quarterly earnings calls.

The amount of this kind of unstructured content is accelerating at an unprecedented rate, making it time consuming to analyze.

As a result, unstructured content is underused as a source of insight. It may contain hints that would quantify a trading strategy, but the overwhelming volume of data makes it impossible to spot the nuances that could drive a decision-making process.

Natural language processing (NLP) offers opportunities to uncover meaningful insights from under-used content.

“NLP is a growing area of artificial intelligence, in part assisted by rapid growth in infrastructure, such as computing power and data handling capacity.

In addition, there have been a number of key algorithmic improvements, and a proliferation of open libraries such as the BERT NLP framework, released by Google in 2018,” explains Rocha.

Efficiency: automating the analysis of volumes of unstructured content in real-time

Speed: the value of the information declines rapidly so insights need to be harvested swiftly

Consistency: a single model achieves consistency that is not achievable if performed by a number of human analysts, each of whom may interpret aspects of text slightly differently

Accuracy: unstructured documents can be lengthy, and human analysts can potentially miss or misinterpret information

A model can be trained to learn how to extract meaning from text, allowing applications and services that understand human language to be developed.

Practical examples of NLP in financial services include speech recognition and intent parsing used by voice assistants and chatbots in customer services, and information retrieval and sentiment analysis of corporate documents and news feeds.

1. Speech recognition and intent parsing
2. Content enrichment - retrieval
3. Content enrichment - trends and relationships
4. Sentiment analysis

Speech recognition is a key piece of the analysis of companies’ quarterly or semi-annual earnings calls.

Corporate conference calls usually start with the company making a presentation on the performance of the previous quarter and the outlook for the following one, followed by a Q&A session in which analysts ask direct and specific questions to the company.

“What and how they ask the questions, and what and how the company answers, including their tone, are likely to reflect on the company’s stock price. Profiling the tone of speech, and converting it to text to quantify it across different key topics, such as revenue, is extremely useful.”

NLP can also be used to retrieve information from unstructured text. This approach is known as named entity recognition (NER), and is used to detect and label entities, that is, real-world concepts, such as people or companies.

NER effectively overlays context on the content by tagging it with machine-readable metadata aligned with an ontology. It’s like having a very detailed Dewey library system, and it means that information retrieval is efficient and accurate.

NLP can also be used to support banks’ compliance processes. Tagging unstructured data facilitates searching across thousands of digital documents, allowing compliance officers to swiftly determine whether regulations have been followed.

“NLP can also be used to create explicit links between supply chain relationships. If the demand for certain products is likely to increase in the near future, then identifying key raw material suppliers would be extremely useful from an investor’s point of view,” adds Rocha.

“Similarly, if a supply chain is expected to be disrupted for some reason, topically, by COVID-19, NER could identify which companies would be affected and to what degree.”

In the investment sphere, applying tags to highlight the main topics covered by text, or topic modeling, is valuable when analyzing earnings calls to establish a main theme, or to compare against previous, similar calls to identify trends.

NER offers additional value, since it can be used to link entities and build a graph of relationships. For example, an entity-modelling system can pick out mentions of specific topics within a range of unstructured text and build new connections.

It can help track relationships between entities, with the potential to detect money laundering or fraud.

Another area of NLP is sentiment analysis, which can extract the subjective meaning from text sufficiently well to be able to determine its attitude, or sentiment. It is an ideal tool for reviewing unstructured content about a particular company to look for inconsistencies and anomalies.

Refinitiv Labs is currently training a new model to identify potential signals of equity performance from thousands of research reports and company transcripts, by identifying changes in outlook over time as potential drivers of equity performance.

“We are currently quantifying the sentiment on more than 100 key drivers of equity performance across different content types: equity research reports, transcripts, news, and company filings,” shares Rocha.

Sentiment analysis can help classify news stories based on positive and negative sentiment to indicate the likely impact on a stock price, but also has more nuanced uses.

“By combining equities’ sentiment scores across different dimensions with a variety of metrics on the evolution of COVID-19, such as the number of cases, death rates, recovery rates, or active cases per capita, we could potentially identify the key drivers of equity performance, including what stocks are affected most, and to what degree.

“NLP could be used to pair COVID-19 mentions in unstructured content to sentiment-based signals. Although the strength of the signal could vary based on geography and industry, having an aggregated sentiment on COVID-19 at the equity level and across different content types could be used to predict future, market-adjusted stock returns.”