In the bustling world of financial services, with continuous market fluctuations, numbers are king.
One investment firm aims to harness the onslaught of qualitative information out there, using AI to wrangle to its advantage a surprising challenger to the field: words.
Toronto-based Triumph Asset Management (recently reorganized as Amadeus Investment Partners) is exploring this area using deep learning. They’re analyzing tens of thousands of news articles each day with the goal of better predicting market directions and making trading decisions.
Old School Style
For many years, groups of human analysts combed through relevant news articles, determined the sentiment towards a specific company, and communicated with traders to react in the market.
The process is time consuming and yields many missed opportunities since only a limited scope of articles can be analyzed, Triumph data scientist Andrew Tan said.
To adapt to a rising need for greater data, the firm turned to AI.
“We believe that with deep learning, utilizing its speed and accuracy, we can improve general analysis and overall workflow about news,” said Tan. “And this, in turn, will yield better results and overall performance.”
AI as Analyst
Using GPUs and the CUDA deep neural network (cuDNN) library, Triumph’s data scientists feed news from a proprietary database into a deep learning system. Trained to parse an article every three milliseconds, the machine churns through hundreds of thousands of articles per day, an endeavor deemed impossible until recently.
The system identifies hundreds of keywords within the articles. An unsupervised learning algorithm, called GloVe, gives each keyword a number value that the rest of the system’s models can then interpret and work with.
The deep learning system ultimately generates three outcomes: It links articles to appropriate stocks and companies; it discerns a sentiment score ranging from positive to neutral to negative for each article; and it assesses the likelihood of the news to impact the market.
During a time when “fake news” manages to permeate through the traditional news cycle, the company’s data scientists use specific keywords and reputable news sources to boost reliability.
While the system is still in preliminary testing, Tan said its initial 76 percent rate of accuracy is encouraging, given that human analysts often disagree on appropriate sentiment scoring themselves.
“The system’s not perfect, but we can build on this,” he said.
Feature image credit: Lorenzo Cafaro