How NLPs are innovating the finance world

By Cristóbal Cortinez, qant Pacífico Research

AI, and more specifically deep learning, have dramatically improved our ability to process unstructured data, such as images and text. New applications go from object detection for self-driving cars and face identification for access to buildings, all the way to programs that are capable of writing new batman comic books:?https://www.youtube.com/watch?v=fn4ArRmzHhQ&ab_channel=iSola7ion (although I’d say some more work is still needed in this area!).

What about finance? When we think about finance, we typically think about an asset’s (or many assets) price, rate, or other derived property as it changes over time. In other words, a timed series of structured data. In these cases, classical timed series methods such as ARIMA, GARCH or their vector counterparts perform roughly as well as deep learning methods but have the added bonus of not being a black box due to the large amount of underlying theory.

This begs the question: Is there any way deep learning could be further leveraged for finance? We will show an example of how a subfield of deep learning called natural language processing (NLP) is being used in the finance world in innovative ways: integrating textual financial information into the forecasting task.

It would be wonderful to include the fast-paced information coming from Twitter. or an informative article about relevant political events into our analysis. The problem is that there is no obvious way to do this: Should we count the number of times a certain word appears in our text and add it as a regression variable? Which words? If we encode text information in this fashion, information about the ordering of the words would be lost. How do we address this?

If we step back for a moment and think about the previous set of questions, we will realize that what we are really looking for is a way to encode a text in an array of numbers of a fixed size, in such a way that we can regress the future values of what we wish to predict over this encoding of the text information (and perhaps a few other values). This is precisely what deep learning models do!

Perhaps the most notable of these is FinBERT: a language model trained on a large corpus of texts and fine-tuned with over 4000 sentences to predict whether the stock of a company will increase or decrease. This model gives us the “sentiment” in a finance-related text that can be directly used in our regression. Instead, we can also go deeper and look at the text encodings that this model used for predicting this “sentiment” and use these instead, probably giving us a better representation of texts and thus a better prediction.

This is only one of the many ways NLP can be leveraged in the field of finance. We should keep an eye open for these kinds of applications, as they are probably key to the future of finance.