This article is based on a presentation at IBA’s Fall 2023 Accounting and Finance Forum: Data with Dignity by Amit Grover, Director of Finance Data Sciences at Discover Financial Services at the time of this presentation.
In just five short years, artificial intelligence has revolutionized the accounting industry. In his role leading the finance and data sciences teams at Discover, Amit Grover has experienced that remarkable transformation firsthand.
Several key drivers have reshaped finance, says Grover, who leads the development and implementation of machine learning models that drive Discover’s business forecasts, allowances for loan losses and leases, and decisions on capital adequacy.
First, the advent of cloud computing has provided on-demand access to computing resources and software. Businesses can build and run complex models without having to invest heavily in expensive hardware and software.
Second, a variety of software tools are now available that can automate many of the functions involved in financial modeling such as data collection, cleansing, and visualization, so that models can be more complex, more accurate, and significantly easier to build and update.
Third, the tools to access big data have become more mainstream and easier to scale—even in small functions that have not traditionally utilized big data.
Fourth, companies are hiring data scientists in finance functions. Traditionally, a financial analyst would have a strong business acumen, accounting skills, and training in tools like Excel. Today, these skills are often complemented by data sciences and business intelligence skills that allow teams to leverage data beyond the numbers to find their business utility in a broader context.
Fifth, the Financial Accounting Standards Board introduced a new accounting standard for estimating current expected credit losses (CECL). This methodology requires financial institutions to include predictive information in their calculations of bad debt.
Predictive Modeling Using AI
Under new CECL standards, financial institutions are required to maintain a certain amount of regulatory capital based on the riskiness of the organization’s assets. That capital is meant to support regular operations and absorb unanticipated losses and declines that could otherwise cause the institution to fail.
For its credit card business, that means Discover needs to be able to accurately predict the amount of debt that they may not recover from its customers.
“Let’s say a customer swipes a credit card to make a purchase,” explains Grover. “Discover bank authorizes the payment and pays the merchant, and the customer has an obligation to pay back Discover. But when there’s a possibility that the customer won’t make the repayment, it creates an allowance for credit loss (ACL) entry. The ACL entry is a contra-asset entry on your balance sheet reflecting uncollectible receivables.”
Anyone who makes a living forecasting finance and economic trends knows that it’s notoriously difficult. And this is where AI can serve as an immensely powerful tool.
“What makes credit cards unique and machine learning models complex to implement is that we have copious amounts of data—millions of customer observations, thousands of attributes per customer, hundreds of economic indicators, and decades of data. But those things can’t be the be-all-end-all of critical decision making.”
Applying Human Judgment to Machine Learning Decisions
“The question I’m often asked is why we need to use judgment,” says Grover. “If the machine has consumed all the data it possibly can, how can a human ever beat the machine? But that’s not the intent. What we’re trying to understand is whether what the machine spits out is reasonable—to understand what the machine knows and does not know. Does the machine have context that humans do? And if not, what compensating controls can we build in?”
In other words, machine learning models are only as good as the data they are trained on and may not function under some scenarios.
For example, after being trained on data from 2020 and 2021 during the Covid-19 pandemic, a model may suggest that high unemployment rates are associated with low levels of credit card losses. But in reality, the CARES Act bolstering replacement income kept many customers afloat and allowed them to continue making their payments despite unemployment. Expert judgment would be needed to account for that context.
Similarly, a model may not account for a change in regulation like the recent resumption of federal student loan payments. Nearly 43 million Americans are resuming an average payment of $390 per month, which they may not have budgeted for and could ultimately impact credit card losses. In this case, expert judgment would be needed for appropriate risk management.
So how should companies manage the balance of machine learning and intervening human judgment?
“My best advice is to rely on a transparent, repeatable and well-supported process,” says Grover.
Select the right experts. Involve stakeholders that understand business processes, and embrace all thought processes.
Leverage empirical data. Look at industry insights and engage appropriate analytical tools.
Prevent imprecise estimates. Produce a range of outcomes and consider all information.
Avoid cognitive bias. Seek multiple perspectives and provide an effective forum for challenge.
“Always start with, ‘What does the model not capture?’” says Grover. “Resolve any inconsistency between accounting and other business practices, ensure appropriate quantitative support for qualitative processes, and put strong governance in place to ensure that models are used and interpreted appropriately.”
Leave a Reply