Cash & Working CapitalFXTechnology

Machine Learning Wizardry: Analytics-Based Cash-Type Forecasting

By September 27, 2023No Comments

Data science “wizards” and clean, treasury-tagged cash transaction data are transforming cash forecasting at one corporate.

Transaction tagging, data science “wizards,” statistical models and machine learning (ML) are the current cornerstones of an ambitious, yearslong project by treasury at one NeuGroup member company to transform cash forecasting—which now includes grouping cash transactions by type. The FX leader and a member of his team shared details of the project, some of its successes and what they have learned along the way at a recent peer group meeting of NeuGroup for Foreign Exchange sponsored by Societe Generale.

  • The move to what the company calls analytics-based cash forecasting is aimed at overhauling the current, manual and resource-intensive process involving more than 150 global entities and forecasters contributing monthly currency cash forecasts used by treasury to produce rolling, 13-month cash flow forecasts for more than two dozen currencies.
  • “These forecasts drive all our US dollar positioning, investments as well as our hedging,” the member said. “When they come in wrong, we’re educating people why it’s such a big deal: We run the risk of needing to change our hedges, offset trades, or miss hedging opportunities.”

Beginning of a journey. The project began with the risk management and cash teams working with a consultant to benchmark with other corporates, evaluate third-party solutions and do a feasibility study on model-driven forecasts. “Let’s not have those 150 forecasters do this each month if we can build a model to do it as well or better,” the member said, describing the goal.

  • That led to a proof of concept initiative using bank account balances, data scientists and ML models. But while initial results were promising, the team ultimately decided that account balance forecasting wasn’t sufficient and went back to the drawing board.
  • “When the account balance forecast wasn’t right, we had no way to answer the treasurer or other stakeholders as to why our 12-month cash forecast was off,” the member said. “We had no real insights.”

Clean data and tagging. To address that problem and identify why and where variances occur, a transition to so-called cash-type forecasting began: all cash transactions—inflows and outflows—would be grouped by type. At the highest level, the types are the corporate’s key drivers of cash flow actuals: payables, receivables, taxes and payroll. (By now, an internal analytics team had replaced the consulting firm and started readjusting the models.)

  • For the project to work, treasury needed to provide the analytics team with clean data and embarked on a massive, global tagging project involving all regional treasury centers. They had to tag all historic transactions and supply business logic for rules-based tagging for future transactions.
  • So far, more than five million transactions have been tagged with some 150 tags representing all transactions. Those 150 transaction types are filtered into about 70 mid-level cash types and about 40 high-level cash types.
  • This split into tiers will help treasury create different levels of forecasts and get more granular to explain variances beyond the main categories of inflows and outflows. Each currency is tagged and treasury can view data by currency alone, cash type by currency as well as by cash type and entity by currency.
  • The FX leader noted that the company had 98% visibility to its transactions by type within its FIS Trax system, giving it a significant leg up. “You have to be able to see that data to be able to tag it properly and know that you have that visibility to feel confident you’re capturing everything,” he said.

The wizards. Equally important to the treasury team’s efforts are those of the data scientists who are on the transformation team of the corporate’s global finance organization. The FX leader’s colleague calls them wizards. “They are a group of wizards who love statistics and are helping us along this journey,” she said. “They are the ones that are building our statistical models and machine learning forecasts. “

  • She explained that these data scientists have built a number of algorithmic models based on statistical time theories, machine learning, neural networks and regressions, among other tools.
  • The models are based on thousands of data points generating thousands of forecasts that are aggregated into one forecast. The models are constantly being refined based on feedback from the risk management team.
  • Treasury also provides the analytics team with information on cyclicality, M&A or anything else that could change the trajectory of the model. “It’s a constant feedback cycle,” the team member said. “The biggest thing we’ve learned with this project is not only does it take a lot of time, but it’s a constant innovation cycle.”
  • Forecasters access the model via an interactive user interface built with Power BI. The model can be refreshed in about 45 minutes.
  • The team is now working to embed internal forward-looking drivers including sales forecasts and business plan information as well as external inputs including economic indicators.

Success and the road ahead. The member’s presentation included accuracy and error data for three of the seven currencies the team has created currency models for thus far. The 12-month error metric ranged from 1% to 13% for the three currencies. The team is also tracking how often the model is meeting the corporate’s internal metric for accuracy. In each of these currencies, the member noted that the analytics-based cash forecast is more accurate than their legacy process.

  • “We’re seeing some nice pockets of opportunity,” the FX leader added. “It’s clearly better in some spaces for us to use this model-based forecast than our existing process.
  • “And it’s really a question of how long it will take us to get all of our currency forecasts built and ready for a transition from our legacy process to the new, machine learning model. We’re thrilled with where we’re seeing it go.”
Justin Jones

Author Justin Jones

More posts by Justin Jones