Brendan Stec
- Nov 14, 2017
- 5 min read

What’s the skinny? Artificial Intelligence and Finance

We’re not talking about Allen Iverson (or practice) anymore. We’re talking about AI technology, and we’ve been hearing about it everywhere: self-driving cars will replace Uber drivers, IBM Watson will work more and more with doctors, and document discovery will increasingly impact lawyers. All mainstream and intuitive applications. But what about AI’s role in financial services?

We can separate AI’s involvement with finance into four main categories:

Front-Office: Applying advanced machine learning algorithms to churn through reams of data, assess credit quality, price insurance contracts, and even interact with clients via chatbots. Tala is a company that uses alternative data (like smartphone data) to assess the credit quality of individuals living in countries where traditional credit scoring is unavailable, providing credit opportunities for millions. Insurtech is an emerging industry that uses AI to customize policies and dynamically price premiums based on constant streams of data.

Back-Office: This goes way beyond replacing Excel macros and reporting. Banks are now using AI to optimize capital allocation - such as executing certain hedging trades to reduce risks and comply with regulations. Machine learning algorithms will be a robust tool for stress-testing and risk management, although it's difficult to tell whether or not they will successfully detect "six sigma events" or black swans . Within trading operations, new "trading robots" will learn from a constant stream of price and volume data to better time trading execution and reduce market impact. In bond markets, clustering algorithms will group together similar securities, so that a trader can model an illiquid security with a substitute that is more liquid and easier to price.

Portfolio Management and Hedge Funds: The main argument here is that AI will quickly see patterns in historic market data to find unique, alpha-generating trades and investments. Kensho's technology allows users to ask in plain English - for example - whether oil stocks have under or outperformed after periods of geopolitical turmoil in Venezuela, which will be a useful tool on the sell-side. Many sell-side firms also have nascent in-house natural language processing (NLP) applications to scan research articles, detect sentiment, and summarize the main ideas so research analysts can succinctly understand the key ideas. On the buy--side, BlackRock is experimenting with using AI to manage portfolios and select securities. Keep in mind, though, certain systematic and quant hedge funds like Renaissance Technologies and Two Sigma have been relying on AI for years now.

RegTech and Fraud Detection: Regulators will use machine learning and piles of bank transfer and accounting data to monitor potential money-laundering activity and the financing of terrorism. Private firms will use the technology to monitor fraud risks, ensure regulations are met, and even flag employees who could be insider trading. The SEC uses machine learning to scan filings and search for potential discrepancies and risks in investment managers. The obvious question here is how much personal data should regulatory agents or employers have access to based on privacy concerns?

AI's increasing role in financial services will have several positives. Firms will benefit from increased efficiency and lower costs, algorithms that better understand customers' needs, and better handling of risks (like potential cyber-attacks or data-quality issues). Customers will enjoy lower fees on different financial products, and thanks to Robo-Advisors and FinTech lending services, will have much easier exposure to them. The entire financial system will benefit from more accurate fraud detection, regulatory compliance, and hopefully improved understanding of the systematic risks in the economy.

Let's be clear, though: AI does have disadvantages and is by no means a "cure all." The first issue here is the lack of interpretability embedded in many machine learning algorithms. While linear and logistic regression do have classic statistical interpretability (coefficients, T-stats, etc.), deep neural networks and other advanced algorithms are "black boxes", meaning users only really see what goes in and what comes out, but nothing about which variables matter. Machine learning algorithms can only make better predictions; they cannot determine causality. If machine learning will be used in risk management, the challenge will be overcoming the black box barrier to find out what is actually causing an increase or decrease in risk.

Another issue with AI relates to a systematic problem in all of finance: the fact that the data the algorithms train on is observational and not experimental data. I cover this issue in more detail here, but the basic idea is this: when looking at historic financial or economic data, we can only observe what happened in the past, and we can't run controlled experiments or quickly gather more data to better tune the model. If Waymo wants to better train its driverless AI, it just needs someone to drive around the parking lot for a few more hours so the model has more data to learn from. But economists can't generate new economic data with the flip of a switch...they need to wait to see what happens! Given this inherent issue and the fact financial data tends to be very noisy/irrational, the concern is the "Garbage In Garbage Out" issue will hurt AI's ability to select securities or manage risks any better than a human can. But employees, policymakers, customers, and the Average Joe may be under the false impression that AI is somehow incapable of failure or mistakes.

I could go on and on about this (problems of spurious correlations, lack of independent time periods, etc.) but want to spare those who aren't exactly thrilled to be reading about the epistemology of AI on a Tuesday morning.

If you're thinking about continuing or starting a career in financial services, what are some points to keep in mind? The big fear is that AI - or GASP... "automation" - will somehow eliminate all jobs. But many experts agree that "augmentation" - the joined forces of AI and human workers - will be the best driver for future economic growth. In wealth management, for example, firms will still need human relationship managers to build trust with clients and handle complex transactions or wealth planning, but will also leverage AI to streamline operations and improve portfolio management. Overall, there will be a continued demand for individuals with high emotional intelligence and the ability to work well with others. Roles that encompass a lot of routine tasks and little interaction with others, on the other hand, are at a high risk of automation. Certain accounting and underwriting roles come to mind here, while venture capitalists and salespeople do not. Regardless of where you work within finance, however, one thing is true: AI is slowly seeping into the industry and will probably be here to stay.

One final thought to conclude with: in 1930, economist John Maynard Keynes predicted his grandchildren would only work about 15 hours a week thanks to advances in industrial technology. We're still waiting John!