The financial sector is entering a new era of rapidly advancing data analytics as deep learning models are adopted into its technology stack. A subset of Artificial Intelligence, deep learning represents a fundamental discontinuity from prior analytical techniques, providing previously unseen predictive powers enabling significant opportunities for efficiency, financial inclusion, and risk mitigation. Broad adoption of deep learning may over time increase uniformity, interconnectedness, and regulatory gaps.
Still in early stages of adoption, deep learning is already being used in finance for fraud detection, regulatory compliance, market surveillance, and administration. It is starting to be used in trading, asset management, risk management, credit underwriting, and insurance underwriting.
Further, through Natural Language Processing (NLP) applications, deep learning is beginning to transform user interfaces, client onboarding, and insurance claims processing. While these applications are not yet truly dominant in finance the way they are in vision or speech, deep learning still comes out on top, after careful tuning, in many tasks. It is likely, even if one assumes only today’s modest benefits, that much broader adoption is yet to come. With advancements in the technology, it is likely that deep learning will gain significant traction in critical finance functions of credit allocation, insurance underwriting, internal risk management, and trading.
AI in Finance
Finance, technology, and data analytics have long existed in symbiosis. Artificial Intelligence (AI) and more specifically deep learning are just the most recent innovations in data analytics to be leveraged by the financial sector. Deep learning builds upon a significant period of transition which brought the internet, mobile phones, cloud computing, and more recently the open banking movement into the financial sector’s technology stack. The introduction of deep learning, with data processing capacity and its predictive prowess, builds on top of and leverages these existing technologies.
Customer interfaces and interactions have been transformed by the deep learning subfield of NLP (Natural Language Processing). It has been key to more intelligent and responsive chatbots and automated call centers, enabling more efficient and possibly more effective customer service. Robo-advisors and virtual assistants have become abundant, using NLP to interview customers, understand their investing preferences, and make trades in the market on their behalf.
These applications are not yet as dominant in finance as they have become in vision recognition or language processing. Having said that, deep learning is still likely to enjoy widespread adoption. After careful tuning in many tasks, it already comes out on top. Investing in the market or playing blackjack against the house - using a tool that helps win 51% of the time can lead to significant profits. In time, with enhancements in computational power and model development, it is likely to demonstrate growing advantages vs. traditional analytics leading. Though much broader adoption is yet to come.
Systemic risk is the risk that events or failures involving one actor, either a firm or individual, or one market sector propagate out to negatively affect the broader financial system and the economy at large. Economies around the globe have witnessed such events when weaknesses in the banking or financial sector spill out to hurt the general public - with millions of bystanders losing their jobs, homes and savings.
Throughout the nineteenth and early twentieth century numerous economic crises emerged from the financial sector. Modern risk management, financial regulation, deposit insurance and central bank backstops have addressed many of the earlier sources of such systemic risk. The basic fundamentals of finance, however, remain - from time to time risks internally built up and concentrated within the financial sector harmfully spill out to the rest of an economy. Most recently we witnessed the devastating ramifications of the 2008 financial crisis, with millions of people losing their jobs and homes, in the USA and around the globe.
The real-world consequences of the periodic crises have motivated rigorous research on systemic risk and underlying firm-level risk sensitivities from both the academic and regulatory communities. Many regulatory organizations around the globe have focused on classifying the attributes of firms that may make their failure more likely to propagate widely.
There also is a very extensive and important body of academic research concerning systemic risk within the financial sector which generally categorizes sources of fragility into one of three primary channels. The first is about uniformity or monocultures, including that which arise from herding. The second relates to interconnectedness. The third concerns the effect of gaps in the regulatory frameworks themselves.
Broad adoption of deep learning in finance is likely to threaten financial stability in meaningful ways. To explore whether the growing maturity of deep learning might awaken systemic risks’ stormy clouds, we accompany the technology along five pathways:
● Model Design
● Algorithmic Coordination
● User Interface
For each pathway, we examine how the nine key characteristics of deep learning may lead to increased systemic risk through underlying firm-level risk sensitivities and the channels of herding, network interconnectedness, and regulatory gaps. We also explore how systemic risks may manifest differently in developing economies with less advanced technology, finance, and regulation.
The key characteristics of deep learning - features of hyper-dimensionality, non-linearity, non-determinism, dynamism, and complexity; challenges of explain ability, bias, and robustness; and an insatiable hunger for data. The advent of deep learning – which combines these nine characteristics together – marks a fundamental discontinuity enabling significant opportunities for enhanced efficiency, financial inclusion, and risk mitigation. Over time, however, broad adoption of deep learning may also increase uniformity, interconnectedness, and regulatory gaps, leaving the financial system more fragile. Existing financial sector regulatory regimes - built in an earlier era of data analytics technology - are likely to fall short in addressing the risks posed by deep learning.