Monitoring borrowers is one of the most important tasks for financial institutions to correctly assess credit risk. This is often done using credit ratings, that have been developed using quantitative methods on financial data. Since key financial data is often updated yearly through annual reports, the question arises of whether threats can be recognized on time. News headlines appear on a much more frequent basis than annual reports, with Bloomberg bureaus alone covering around 10000 news messages daily.

Keeping in mind that news must be carefully filtered from noise, the use of news headlines sentiment in a credit risk setting is an example of how unstructured data can be used. The tone of news headlines says a lot about how investors perceive the market. The field of behavioral finance shows us how important this perception can be in the decision-making process of investors. They often act upon perceptions gained from consuming media. This makes news headlines incredibly powerful, as they not only influence our opinion, but also indirectly influence reality itself. Sentiment can therefore be described as the way that the information is perceived by the market. A timely assessment of this sentiment is useful for pricing financial instruments, a better understanding of the state of the economy and of the risk associated with a counterparty. The rise in interest in NLP models of the past decade plays an important role in retrieving a sentiment value, as the more sophisticated models can interpret complicated news headlines.

Measuring News Sentiment with NLP Methods

One of the most traditional approaches in extracting sentiment from text is the lexicon-based method (LM). In this method, a labeled dictionary with positive and negative words is applied to a snippet of text and the sum of those values classifies this snippet as positive or negative. In a financial context, such a dictionary is provided by Loughran and McDonald (2011) and contains more than 4000 labeled words constructed from a financial point of view.

Although the method described above is simple and easy to use, its disadvantage is that it does not capture the full nuance of words. Linguistics is a very complicated subject that takes the human brain years to master. The most intuitive example is the fact that the words 'dog' and 'cat' would be considered as two completely different words within the lexicon method, whereas we would like to at least consider the fact that they are both animals and are in that sense similar to each other. Deep learning methods can help in understanding this with the use of so-called embeddings. These are representations for text where words that have the same meaning have a similar representation. Combining this with an NLP method such as Google's BERT (Bidirectional Encoder Representations from Transformers), one can combine this broader knowledge of the language with sentiment weights for sentences so that the model learns when to interpret a certain sentence as positive and when as negative, using neural networks. A neural network is a series of algorithms that endeavors to recognize underlying relationships in a dataset and mimics the functions of a human brain. Neural networks are even more powerful in computational terms than the human brain and can learn a language in half an hour rather than taking many years.

News Sentiment as Early Warning System

News sentiment is already being used a lot in stock price forecasting. Studies have shown that there exists some correlation between the sentiment of de-noised news and the direction of stock prices. Compared to equity markets, the use of NLP within risk management is relatively new. A lot of financial institutions focus on a stable outlook of creditworthiness of a counterparty using so called Through-The-Cycle models. As these models do not fluctuate according to current changes in macro-economic circumstances, they provide a long-term outlook of creditworthiness. The main drawback of these models is that they are infrequently updated. If a company defaults during a year where the credit rating indicated a low probability of default, the question arises if such an event could be foreseen. Forward looking models could give an indication of the current fluctuations in creditworthiness, however it is difficult to assess if these fluctuations are rational or caused by irrational behavior in the market. An early warning system that could interpret current fluctuations and indicate where these fluctuations originate could reduce the number of unforeseen defaults.