Cash Flow

Machine Learning Advances the Cash Forecast

Treasurers assess what machine learning can and cannot yet do for cash forecasting.
Machine Learning Advances the Cash Forecast
Photo: Getty Images

“One of the most important things that treasury can do is to come up with a good cash forecast,” Bruce Lynn, managing partner of The Financial Executives Consulting Group (FECG), told CFO last April. 

Especially in the current economy. Treasurers would rather not over-borrow, under-invest, or be unduly exposed to high-interest rates and volatile foreign exchange rates, he explained.

Accurate information on the cash position and future cash flows can prevent those costly mistakes.

That vendors and banks are introducing artificial intelligence and machine learning capabilities into cash forecasting is no secret. The question is how much these technologies will help companies produce and improve the notoriously onerous and sometimes frustrating task of cash forecasting. 

Predictive Applications

Applying machine learning (ML) to cash forecasting makes a lot of sense. Machine learning problems typically involve predicting previously observed outcomes using past data. “The technology is best suited to solve problems that require unbiased analysis of numerous quantified factors in order to generate an outcome,” wrote Chandu Chilakapati and Devin Rochford of Alvarez & Marsal in CFO three years ago.

But ML, in its current state, is not a magic black box. Treasury departments may be using ML in forecasting, but lots of human intervention and experimentation are required.

What AI can do is “boost analytics and decision-making capabilities and deliver information and insight at the right step in the workflow,” said Tommy Wimmer, head of data & analytics solutions at JP Morgan Chase, at the Association for Financial Professionals (AFP) annual conference last week.

It’s given me a third data point to put in front of our CFO and other members of the leadership team to give them an idea of what better to expect. — Jordan McFarland, senior treasury analyst, Evoqua Water Technologies

In a session titled, “Using Machine Learning to Build an (Artificial) Intelligent Forecast,” Wimmer revealed that JP Morgan has more than 150 data scientists and engineers working with payment-flow data to refine its ML forecasting solution. (It is not the only bank, however, that has added these technologies to its cash forecasting tool.)

Some of JP Morgan’s treasury clients use their banking transaction data to forecast cash to either supplement or compare with data coming out of accounting or enterprise resource planning systems.

Panelist Tara Ashmore, director of corporate finance for Amtrak, produces a 13-week cash forecast weekly and shares it with two companies that manage the national passenger railroad operator’s $3.5 billion in investments. The forecast helps them “better time [Amtrak’s] liquidity needs with investments, so we’re maximizing interest income,” she said.

Ashmore likes the JP Morgan ML features because she can get good predictive information on credit card receivables, station cash (when passengers pay by cash), and state partnership reimbursements based on prior transaction history. And she can do that without relying on the sales expectations of people within the organization, which may be biased.

But the cash forecasts aren’t seamless. Amtrak gets federal government funding in “big chunks” once or twice a year, said Ashmore, so that bucket of less-predictable receivables has to be removed, otherwise it would skew the forecast. 

Telling the machine not to take into account “outlier” transactions when it produces the forecast “has been one of the trickiest elements of cash-flow forecasting to build in the software,” Wimmer said.

Jordan McFarland, senior treasury analyst at Evoqua Water Technologies, has two different cash forecasts, both done in Excel and highly manual. One involves inputs from five different members of the organization, he said. “I’m spending a lot of time interpreting the data I received and finding a way to move it into my master workbook that consolidates the data,” he told the AFP audience. 

Running JP Morgan’s tool is easy, McFarland said, and allows him and the treasury staff to spend more time analyzing the forecast.

But the enhanced cash forecasting tool from JP Morgan hasn’t replaced those Excel-based forecasts. “It’s given me a third data point to put in front of our CFO and other members of the leadership team to give them an idea of what better to expect,” McFarland said. “There’s no harm in adding more to your arsenal.”

Data Needs

How would a treasurer or other finance executive get started using ML-augmented cash forecasting, an audience member asked. The algorithms probably work best if the company has two years of data with the bank, the panelists said. 

But a new bank may be willing to ingest data from a company’s current system to establish the history, “whether you were with them from day one or not,” said panelist Frank Woodley, director of treasury operations at Arcosa, a product and service provider to construction, engineered structures, and transportation markets.

We believe in a world where at some point AI will execute certain simple tasks and find patterns invisible to the naked eye but we’re not quite there yet. — Tommy Wimmer, head of data & analytics solutions, JP Morgan Chase

Said Wimmer: “Even if you only have a year or half a year, you can start to see patterns over the quarters and can really drill into the past.”

At the extreme, only a month’s worth of cash data can be valuable with ML-augmented tools, said Amtrak’s Ashmore. An example would be a quickly observable trend like the peak in payables the last week of a month, particularly the last two days of the month, she said.

“And you can share that with the CFO and say, ‘You know this happens all the time. Why can’t people put their invoices in more regularly?’ So it will prove useful in ways you might not have imagined.”

As to whether an ML-augmented forecast is more accurate, Evoqua Water’s McFarland said forecasting should be thought of as a journey. “In the first iteration, [the forecast may be off] a large percentage. But that [leads you] to investigate further and see what assumptions you made,” he said. “And then you tweak your model or tweak your approach and in the next quarter hone in a little bit closer. You’re continuously improving based on what you’re learning about the business each time.”

Currently, the AI capabilities make predictions that supplement human expertise, Wimmer said, in agreement. “Where this is headed” and where JP Morgan is investing is in the system making recommendations, he said.

For example, the tool would one day be able to flag a potential cash-flow shortfall in a region and suggest moving cash to a bank or account tied to that geography. “Then the human can say, “I liked this recommendation or, no, this recommendation is completely wrong,” Wimmer said. The treasurer will set boundaries “within which AI can execute basic tasks on their behalf,” said Wimmer.

“We believe in a world where at some point AI will execute certain simple tasks and find patterns invisible to the naked eye,” said Wimmer, but “we’re not quite there yet.”

The journey to optimal use of ML in cash forecasting is a four-step journey, he added, “and we’re at about step two-and-a-half.”