Why AI cash forecasts are only as good as the data behind them
AI can make cash forecasting more adaptive and dynamic—but only if the underlying data is reliable.
Cash forecasting is shaped by context and technique. Models built on historical patterns tend to perform well when conditions are stable, but they are tested when volatility is persistent. Across finance functions, the question is how quickly forecasts adapt when assumptions change.
For Philomel Peña, Treasurer at University of California, Office of the President, this environment explains the growing interest in artificial intelligence for cash forecasting—not to replace judgment, but to make forecasting frameworks more responsive. AI offers the possibility to “really shift it from being a very static kind of assumption-driven type of exercise to something far more dynamic and adaptive.”
That shift is partly about reallocating effort. Treasury teams spend significant time maintaining models rather than interpreting outcomes. As Peña puts it, the aim is to “utilise our own resources… the human resources to really utilise their time most efficiently instead of relying on and inputting static data.”
Much of that manual effort goes into managing seasonality and cyclicality—tasks that assume stable patterns. In practice, they rarely do. Machine learning, she expects, will automate that adjustment: “the seasonality and the cyclicality of things will be more automated… as opposed to us having to make those changes manually.”
The limitations of current tools are clear. “Our current models are really very manually driven,” Peña says. Machine learning, by contrast, can “detect real structural differences or changes and evolving seasonality without requiring… a human to identify each and every one of it.”
As payment behaviour, counterparty timing and volumes shift, machines can “respond a lot faster” than human analysts scanning for incremental changes.
AI’s real strength lies in detection rather than prediction. It can identify “nonlinear and multifactor interactions where we may look at something more linear.” For treasury teams—“typically… the smallest team in the finance organisation”—that matters.
AI also changes how uncertainty is handled. Instead of a single forecast, AI supports probabilistic thinking. Peña describes a move towards “a distribution of what those forecasts would look like,” allowing teams to “quantify uncertainty and assess… tail risk a little bit more effectively.”
Scenario modelling becomes “more dynamic,” incorporating a wider range of outcomes.
Still, the limits are clear. AI “does not forecast spontaneity,” she notes. Nor can it interpret intent. One-off events should not be smoothed: “you don’t smooth a one-time event, right?” Strategic decisions, regulatory changes and future policies also remain invisible to algorithms.
Most importantly, AI amplifies whatever data it is fed. “Garbage in, garbage out,” Peña says bluntly. Forecasts, she adds, “will be just as good as your data.” Implementation often exposes deeper flaws, including banking structures “created way before we got here” that no longer serve today’s needs.
AI will not eliminate uncertainty from cash forecasting. But it may finally force models to change when the world does—rather than pretending it hasn’t.
Philomel Peña, will be speaking at the 11th EuroFinance Annual Treasury Cash
Management Summit West Coast in the session “Cash management & forecasting with AI” alongside Bridget Rodnick, Treasurer at Ultragenyx and Garima Thakur, Treasurer at CAA.