Why Moody's is betting on machine learningWhy Moody's is betting on machine learning
The organization is solving the challenges of financial risk modeling with AI
December 10, 2019
The organization is solving the challenges of financial risk modeling with AI
You might know Moody’s Corporation for its credit rating agency, Moody’s Investors Service, trusted by bond investors for opinions on credit risk. But the company, founded way back in 1909, grew ever-larger, and in 2007, Moody’s Analytics was established to focus on non-rating activities - including economic research, consulting services and software development. Today, the employees of Moody's Analytics include a large number of machine learning and deep learning experts.
To find out more about the business, we quizzed one such expert, Ashit Talukder, head of machine learning at Moody’s Analytics, who previously spent 12 years at NASA’s Jet Propulsion Laboratory and served as the CTO of the US Department of Labor.
Q: Most people know Moody’s as a credit rating agency. What is Moody's Analytics?
AT: Moody’s Analytics provides financial intelligence and analytical tools to help business leaders make better, faster decisions. The machine learning team is based in the Moody’s Analytics Accelerator, the firm’s innovation unit. The Accelerator explores business opportunities adjacent to Moody’s core business and rapidly develops prototypes of new solutions by leveraging emerging technology.
Q: How does Moody’s Analytics use machine learning technologies?
Moody’s Analytics has been using technology for analytical purposes for many years. We pioneered the use of statistical analytics models for risk assessment. More recently, we began strategically leveraging machine learning and artificial intelligence across our solutions.
An early focus area was to use artificial intelligence and machine learning to extract knowledge and structured information from unstructured data, such as text, images, video and audio. We observed that unstructured data analysis in finance (as it is in many other domains) is mostly a manual process. These highly manually processes result in costly, ineffective solutions and workflows while limiting the amount of data that can be analyzed and consumed. The result is often that critical information is overlooked due to the finite number of person-hours available. Recognizing that fundamental issue has driven many of our AI and ML initiatives. In other words, we are using AI and ML at Moody’s Analytics to enable faster, data-driven decisions at scale.
To that end, we have launched intelligent products that augment the role of the analyst in the data gathering or data normalization process. At AI Summit 2018 we launched QUIQspread, an intelligent, automated, financial spreading tool that standardizes customer information from income statements, balance sheets and other financial documents
Additionally, we offer Compliance Catalyst with Adverse Media which supports compliance risk assessment in client onboarding, Know Your Customer (KYC) and enhanced due-diligence activities. We have also started to use machine learning for better financial risk modeling and default prediction from financial statements.
Q: How difficult is it to extract information from unstructured data sources, like social media posts or video? Why is this information valuable?
Machine learning has been used for analytics and decision-making from structured data for over two decades. By structured data, I am referring to data in machine readable, row-column formats, in the form of mostly numerical and/or categorical data. Traditional machine learning models (such as logistic regression, decision-trees, SVMs, etc.) are well suited to process structured data. Until recently, the ability of ML to process unstructured data, such as text, images was very limited.
With the advent of deep learning, AI ML models can now effectively process raw unstructured data, without requiring manually crafted features (which are often only partly optimal). Now, deep learning Convolutional Neural Networks can process image data for image classification and object identification; Recurrent Neural Networks and Long-Short-term Memory (LSTM) neural networks can do natural language processing (NLP) tasks by “reading text passages” and using context to understand sentiment, identify key entities and relations between them.
Now it is easier than ever to extract information from unstructured data sources but this is still a problem that is not fully solved. First, text information from social media and news is voluminous, complex, very noisy, with a high degree of variability in expression, and its nontrivial to fully understand context due to the high degree of variability and nuances in human language. So, while Natural Language Processing (NLP) has improved significantly, more robust Natural Language Understanding (NLU) has still a way to go in terms of comprehension and generalizability across domains and applications. So, in short, its actually still a significant endeavor to extract and comprehend information from unstructured sources.
The information available in unstructured data, such as news, company reports, transcripts of meetings and announcements, and social media often complements the information available in structured data (such as numerical data in financial statements). Furthermore, the information from unstructured data is more recent, and often offers real-time insights into events, changes related to organizations and people, that could provide timely insights into early warning indicators and risk models, thereby enabling faster insights and predictions than was previously possible with more traditional data. This, of course, assumes that the AI models can extract the key nuggets of relevant information from the available massive, diverse data sources in a timely manner. Our AI-driven solutions and products demonstrate that with the right models, and training data, we can find the “needle in the haystack” for obtaining valuable, timely insights from unstructured data to better understand and assess different types of risk.
Q: In your career, you spent two years working as CTO of the US Department of Labor. How do you rate the understanding of machine learning in the public sector?
The public sector has traditionally been ahead of the curve in terms of R&D of emerging technologies. Federal agencies such as DARPA, ARPA, NSF, NIST, NASA, NIH have been spearheading investments in basic and R&D around new technologies for decades, such as networking and the internet (the ARPAnet funded by DARPA ultimately led to the invention of the internet), biotech, healthcare, robotics, AI, and machine learning, and more recently quantum computing. NIST has had a Quantum computing program for more than a decade, while industry has only recently used Quantum computing. DARPA initiated robotics and AI research, through R&D investments in the 90’s and 2000-2005 timeframes, through programs in the 1998-2004 timeframe such as Software Defined Robotics (SDR), Robotics Vision 2020, the DARPA Grand Challenge that led to the transition of autonomous vehicles, and AI-enabled computer vision into industry.
Recently, the private sector has led the use of Machine Learning and AI in operational settings (i.e. beyond R&D), once the basic R&D issues in AI, ML were resolved through federal funding and deep learning proved its mettle on research problems. Part of the reason for this is that the private sector is more risk-tolerant in application of emerging technology to solve new problems and faces fewer regulations. Secondly, the public sector deals with legacy systems and non-digital processes that present a hurdle to adoption of AI. This issue is less prevalent in the private sector, where Web 2.0 (websites that rely on user-generated content) companies employing modern systems and digital data, already exist. Additionally, the use of AI, ML in commercial settings is broader, in areas such as retail and e-commerce (recommendation systems), healthcare (insurance modeling and prediction), finance (risk analytics and prediction), and others.
Now with a better understanding of the prevalence and impact of AI, the U.S. government has realized that policies that take ethical considerations into account need to be put in place to accommodate a future where AI plays a central role in shaping the future of society, industry and science. New strategic planning committees are being established in the public sector to help guide policies and investments in an AI-driven world.
Q: You also spent more than a decade at NASA’s Jet Propulsion Laboratory. What are the key applications of machine learning in the aerospace industry?
I thoroughly enjoyed my time at NASA’s Jet Propulsion Laboratory. The aerospace industry and the space research agencies (NASA and its subsidiaries) have been employing machine learning in space and aerospace applications since the 90’s. JPL had an AI group that developed novel rule-based expert systems for spacecraft resource allocation, and a computer vision group and ML group that designed computer vision and ML systems for robotic planetary navigation and exploration. At JPL, through funding from the DARPA Robotics Vision 2020 Research program, I co-led the development of the first real-time computer vision solution capable of detecting and tracking moving objects from moving autonomous robots in 2002; this body of work has been cited more than 850 times. We also worked on distributed ML solutions for IoT earth observing sensor networks in the 2005 timeframe. A lot of R&D work around autonomous systems and AI at NASA and aerospace agencies was patented, and spun-off into the commercial sector.
Q: Moody’s Analytics is currently expanding its data science team. How hard is it to find the people with the skills the organization requires?
We are always on the lookout for talented data scientists! Given the potential for AI, ML, many organizations are vying for the unique combination of talent that can both understand and implement custom AI ML solutions for practical applications. We are particularly looking for individuals that have a thorough grasp of the theoretical aspects of deep learning, but also understand how the implementation of AI solutions works in practice and have the data engineering skills needed to roll out custom applications.
Deep learning is still nascent and in its infancy. However, my team, based in the Moody’s Analytics Accelerator, offers a unique environment where we span applied AI, ML R&D and product development experience, and the opportunity to spread one’s wings in terms of understanding business and product lifecycles while working on exciting, cutting-edge AI technologies.
About the Author(s)
You May Also Like