Your Forecast Is Only as Defensible as the Data Behind It

TL;DR

Most cash forecasts fail on data, not models. EuroFinance's 2026 research found 51% of treasury professionals name data quality and consistency as the single biggest constraint on forecast accuracy, and 37% distrust AI-generated forecasts because of poor underlying data. Explainable AI only matters once the inputs are explainable: every number traceable to its source, bank data reconciled continuously, and forecast-vs-actual variance treated as a data quality signal first.

Your Forecast Is Only as Defensible as the Data Behind It

Most cash forecasts don't fail in the model, they fail in the data behind it. A treasurer walks into a leadership meeting with a rolling 13 week forecast that took days to assemble. The CFO looks at last month's forecast next to the actual, points at the gap, and asks a version of the same question every CFO eventually asks. Why did this number move so much, and where did it come from?

If the honest answer involves five spreadsheets and a manual VLOOKUP against a bank portal export the forecast has already lost the room. It doesn't matter how sophisticated the underlying model is, a number nobody can trace back to its source is not a forecast the business can act on. It's a guess with better formatting.

Data Problems, Not Modeling Shortfalls

Treasury teams have spent the last few years pointed at the wrong end of the forecasting problem. The instinct when forecast vs actual variance is too wide is to reach for a better model, a smarter algorithm, or now, an AI layer that promises to spot patterns a human would miss. But a model built on inconsistent, incomplete, or unreconciled data will simply produce a more confident wrong answer.

That's borne out in the numbers. In EuroFinance's 2026 Deep Dive on AI in treasury, produced with Treasury Intelligence Solutions, 51% of treasury professionals identified data quality and consistency as the single biggest constraint on forecast accuracy, ahead of every other factor surveyed. That's the majority explanation for why forecasts stumble.

The same research found something sharper when it isolated AI specifically. Asked what makes them distrust AI generated forecasts, 37% of respondents named poor underlying data as their top reason, well ahead of concerns about the model itself, governance, or explainability as a standalone issue. Treasury teams aren't primarily worried that the algorithm is a black box, but that it's a confident black box built on the same fragmented inputs that already made their manual forecasts unreliable.

Explainability Starts With The Data Trail

There's a version of 'explainable AI' that treasury teams are being sold that starts and ends at the model layer. Show your work, surface a confidence score, let the treasurer see which variables the algorithm weighted most heavily. That's useful, but it answers the wrong question first.

Before a treasurer can explain what an AI model did with the data, they need to be able to explain the data itself. Where did this cash position come from? Was it the available balance or the booked balance? What date was it valued as of, and does that match the date the forecast assumes? If those questions can't be answered in the room, an explainable model sitting on top of unexplainable inputs doesn't actually solve the trust problem, it just moves it one layer down.

This is where forecast vs actual variance analysis becomes a genuinely useful discipline rather than a monthly reporting chore, as every material variance is a data quality signal. A recurring gap between forecast and actual in a specific entity or bank relationship usually traces back to a specific normalization problem, not a forecasting assumption that needs tweaking. Treating variance analysis as a debugging tool for the underlying data, rather than just a scorecard for the forecasting team, is what separates treasury functions that can defend a number from those that can only present one.

Most explainable AI finance vendors are selling the wrong half of the problem. The pitch is usually about model transparency, a way to interrogate why the algorithm produced the number it did. That's a reasonable ask once the inputs are trustworthy, but treasury teams asking for explainable AI right now are, more often than not, actually asking for explainable data. They want to know why a bank statement shows one balance while the general ledger shows another, why a wire that settled on Friday shows up in Monday's feed, or why the same account produces a different available balance depending on which portal pulled it. An AI layer that explains its own weighting scheme without being able to answer any of that hasn't solved the trust problem treasury actually has.

The same gap shows up in ordinary treasury reporting long before AI enters the conversation. A daily cash position report that can't be reconciled to the bank statement it's supposedly built from is just as indefensible as a forecast built the same way, and for the same reason. Reporting and forecasting are downstream of the same data pipeline, so a weakness in one is rarely isolated to just the one.

Treasury is Already Moving Ahead

Despite the trust gap, treasury teams aren't waiting for perfect data before exploring AI. The same EuroFinance research found 46% of treasury teams are actively evaluating AI solutions for forecasting, compared with 32% evaluating a new treasury management system outright. Teams increasingly see AI as the more urgent investment, even ahead of a broader TMS overhaul, because the pressure to produce a defensible number is more immediate than the pressure to replace a platform.

That instinct is reasonable, but it inverts the order of operations that actually works. AI evaluated on top of fragmented, unreconciled bank data doesn't fix the trust problem 37% of respondents named. It amplifies it, because now there's a model making judgment calls on inputs nobody has fully validated.

The 2025 AFP Treasury Benchmarking Survey backs this up from a different angle. Over 60% of treasury professionals still name cash and liquidity forecasting as the single most challenging task they face, a number that hasn't moved much even as AI adoption conversations have accelerated. New tooling isn't closing the gap on its own.

Where AI does show a clear, near term return is further down the stack. Around 30% of respondents in the EuroFinance survey pointed to reducing manual reconciliation effort as AI's biggest potential contribution, not forecasting itself. That's showing that the highest value AI use case in treasury right now may not be the forecast at all, but the unglamorous work of getting bank data clean, normalized, and reconciled daily, so that whatever forecasting method sits on top of it, human or AI assisted, is working from a defensible base.

What a Defensible Forecast Requires

A forecast survives the CFO's questions when three things are true.

1. Every number in it can be traced back to a specific source system and a specific point in time, not reconstructed from memory or a spreadsheet formula nobody remembers writing.

2. The reconciliation between bank data and internal records happens continuously, not once a month when the close forces it. A forecast built on a bank balance that's three weeks out of reconciliation is a forecast built on a number nobody has actually verified.

3. Variance between forecast and actual gets investigated as a data question first and a modelling question second, because in most cases that's genuinely where the answer lives.

None of that requires exotic technology, but it does require bank data that's normalized and reconciled automatically and close enough to real time that treasury isn't forecasting off last week's picture of cash. Trovata's approach to daily bank reconciliation exists for exactly this reason, because it's the layer that determines whether anything built on top of it, forecast or otherwise, can be trusted. The teams building a genuine ROI case for forecast automation tend to find the real return sits here too, in the hours no longer spent chasing down why a number doesn't tie out.

The Data Foundation Is The Real Test

It's tempting to treat forecast accuracy as a purely technical problem, one more model iteration or one more AI feature away from being solved. But the treasurers who consistently defend their numbers in front of a CFO are the ones who can answer the follow up question without hesitation, because they trust the data underneath the number as much as the number itself.

Until data quality is treated as the forecasting problem rather than a prerequisite to it, no amount of modelling sophistication, AI assisted or otherwise, will make a forecast survive the room.

Trovata shows you what a defensible forecast really looks like, normalizing and reconciling bank data automatically, so every number in your forecast traces back to a source you can point to. Book a demo today.

FAQs

Why do cash forecasts miss?
Most misses trace back to data problems, not the forecasting model — inconsistent, incomplete, or unreconciled inputs. In EuroFinance's 2026 Deep Dive on AI in treasury, 51% of treasury professionals named data quality and consistency as the single biggest constraint on forecast accuracy, ahead of every other factor surveyed.

What is explainable AI in finance?
Explainable AI in finance means a finance or treasury team can see how a model produced a number — the variables it weighted, the data it used, and the confidence behind the output. For treasury, real explainability starts below the model: if the underlying bank data can't be traced and reconciled, an explainable model can't be trusted either.

Why don't treasury teams trust AI-generated forecasts?
Poor underlying data is the top reason, with 37% of respondents in EuroFinance's 2026 research naming it ahead of concerns about the model, governance, or explainability. Teams aren't worried the algorithm is a black box so much as a confident black box built on the same fragmented inputs that made manual forecasts unreliable.

What makes a cash forecast defensible?
A defensible forecast has three properties: every number traces to a specific source system at a specific point in time, bank data is reconciled continuously rather than at month-end, and forecast-vs-actual variance is investigated as a data question before a modeling question.

Should treasury adopt AI for forecasting before fixing data quality?
No. AI evaluated on top of fragmented, unreconciled bank data amplifies the trust problem instead of fixing it. The clearest near-term AI return sits further down the stack: around 30% of treasury professionals point to reducing manual reconciliation effort as AI's biggest contribution, which builds the clean data foundation forecasting depends on.

How does forecast vs actual variance analysis improve data quality?
Every material variance is a data quality signal. A recurring gap in a specific entity or bank relationship usually traces to a normalization problem, not a forecasting assumption, so treating variance analysis as a debugging tool for the underlying data separates teams that can defend a number from teams that can only present one.