The AI That Barclays and Citi Now Trust to Read Currency Flows

A forecasting engine moves from the leaderboard to the trading desk#

Most artificial-intelligence stories in banking end at the demo. This one did not. In the space of a week, Ant International, the Singapore-headquartered payments arm spun out of China's Ant Group, went from claiming a benchmark record to confirming that four of the world's larger banks have wired its forecasting model into the systems that price and hedge real money. That short journey, from research claim to production use inside regulated institutions, is what makes the announcement worth a careful read rather than a quick skim.

What happened#

On 19 August, Ant International unveiled FalconTST 2.0, the latest version of its Falcon Time-Series Transformer, and said the model had reached state-of-the-art performance on the Mean Absolute Scaled Error (MASE) metric on a leading public benchmark for time-series foundation models. Days later, FinTech Futures reported that Barclays, Citi, Deutsche Bank and Standard Chartered had integrated the model into their own foreign-exchange platforms to sharpen cash-flow forecasting and currency-risk management.

The model's job is narrow but valuable: predict how much money will move, when, and in which currencies. Ant first ran FalconTST internally to manage its own liquidity and FX exposure on an hourly, daily and weekly basis before the banks adopted it. Each institution uses it inside existing infrastructure, as the table below sets out.

Sources: Ant International / Business Wire; TechNode.

Standard Chartered's use runs through the Monetary Authority of Singapore's PathFin.ai programme, a supervised sandbox for AI development in finance, which gives the deployment a regulatory frame rather than leaving it purely commercial.

What a time-series transformer actually is#

To see why banks care, it helps to separate two families of AI. Large language models, the technology behind chatbots, learn patterns in text. A time-series transformer learns patterns in sequences of numbers ordered through time: transaction volumes, account balances, settlement flows and currency positions. The "transformer" is the same underlying neural-network design that powers language models, but pointed at numerical history instead of words.

The newer idea is the foundation model for time series. Traditionally, a treasury team built one model for liquidity, an airline built another for passenger demand, and a retailer built a third for sales. Each was trained from scratch on its own data. A foundation model is instead pre-trained on enormous quantities of time-series data from many industries at once, on the wager that cycles, trends, seasonality and sudden shocks rear their heads in similar mathematical shapes whether the underlying series is jet-fuel demand or weekend card payments. Once trained, such a model can forecast a series it has never seen before without task-specific retraining, an ability researchers call zero-shot forecasting. Ant says FalconTST does exactly this, learning shared temporal structure across finance, retail, energy, travel and economics through a component it calls ORBIT.

That matters for a payments business because accurate forecasts feed directly into capital efficiency. Knowing precisely when funds are needed, how much, and in which currencies, determines how much cash sits idle and how much hedging a firm buys.

Market implications: why one forecasting model touches many markets#

The reach of this story comes from the market it sits inside. Global foreign-exchange turnover ran at roughly $9.6 trillion a day in April 2022, according to the Bank for International Settlements' most recent Triennial Survey, and the value of cross-border payments is projected to pass $250 trillion by 2027. Small improvements in forecasting, applied at that scale, translate into meaningful sums.

In foreign exchange, better prediction of incoming and outgoing currency reduces the twin costs of hedging: over-hedge and you pay for protection you did not need, under-hedge and you carry exposure you did not want. Consider an airline that collects fares in dozens of currencies while paying leases, fuel and airport fees in others. The tighter its forecast, the cheaper and cleaner its hedging book. The same logic extends to fixed income and money markets, where liquidity forecasting shapes how much short-dated funding a treasury must raise, and to institutional investing, where quantitative desks already treat forecasting quality as a source of edge.

For fintech and banking, the more consequential signal is competitive. A payments company is selling forecasting intelligence to the very banks that are also its counterparties and partners, which blurs the old line between fintech challenger and bank supplier. For regulators, an AI model steering hedging decisions across several systemically important banks raises familiar questions about model risk, concentration and what happens if many institutions lean on correlated predictions at the same moment.

Under the bonnet of FalconTST 2.0#

The headline number is a MASE score of 0.666. Mean Absolute Scaled Error, proposed by statisticians Rob Hyndman and Anne Koehler in 2006, measures a forecast's average error against a naive benchmark that simply repeats the last observed value from the equivalent prior period. A MASE below 1 means the model beats that naive rule; a score of 0.666 implies errors roughly a third smaller than the benchmark. The metric's appeal is that it is scale-independent, so results can be compared across series measured in very different units, which is precisely why leaderboards for general forecasting models rely on it.

The leading public benchmark for these models is GIFT-Eval, released by Salesforce, which scores models across 97 task configurations drawn from 55 datasets and reports normalised MASE and a probabilistic accuracy measure called CRPS. Ant says FalconTST 2.0 sits at the top of such a leaderboard, ahead of models from other technology firms, though the company's benchmark claims rest on its own testing rather than an independent audit.

Ant flags three engineering choices in version 2.0. First, smarter handling of missing data: an absence of weekend bank transactions does not mean demand fell to zero, and the model is built to tell a genuine zero apart from a gap so it does not learn a false lull. Second, cross-domain generalisation through ORBIT, letting the model forecast a new business scenario directly. Third, native support for multiple frequencies within one architecture, from second-by-second payment data to hourly treasury positions, daily airline demand and monthly economic indicators. Ant has also opened an API trial on GitHub, which lets outside developers probe the model rather than take the marketing at face value.

Critical analysis: the case for scepticism#

Named, systemically important banks putting a model into production is a higher bar than a paper or a press release, and publishing an accessible API invites the independent scrutiny that vendor claims usually dodge.

The limitations deserve equal weight. The benchmark record is self-reported, and a single aggregate score can flatter a model that is uneven across the sudden regime shifts, such as a rate surprise or a currency peg breaking, that matter most in markets. The reported "forecast accuracy rate consistently above 93%" is not clearly defined and cannot be checked against a stated methodology, so it should be read as a vendor claim rather than an audited result. Foundation models are also known to struggle at long horizons, where recursive one-step-ahead prediction lets small errors compound.

There are structural risks too. If several banks hedge from correlated forecasts, they may crowd into the same positions and amplify moves rather than dampen them, a variant of the model-monoculture problem regulators already watch. Adoption barriers are practical: FX desks are conservative, model-risk governance is slow, and a forecast is only as good as the decisions built on top of it. A more accurate number that no one trusts enough to act on changes little.

From ARIMA to foundation models#

Statistical forecasting in finance is not new. For decades the workhorses were models such as ARIMA and exponential smoothing, hand-tuned to one series at a time. Machine-learning methods then improved accuracy but still, for the most part, learned a fresh model for each task. The shift now underway echoes what happened in language: general-purpose, pre-trained models that transfer across problems. Google's TimesFM, Amazon's Chronos and Salesforce's Moirai opened this line of work; FalconTST is a payments company's entry into the same race, distinguished mainly by its move straight into live bank use.

Whether this is a paradigm shift or an incremental step is the honest open question. The modelling approach is genuinely new relative to the single-task tools banks still run, which argues for structural change. Yet forecasting has been improving steadily for years, and one vendor's leaderboard position does not by itself rewire an industry. The more durable signal is institutional: banks treating a third-party foundation model as core forecasting infrastructure, rather than building everything in-house, points to a lasting change in who supplies financial intelligence.

Key takeaways#

1. Four global banks (Barclays, Citi, Deutsche Bank and Standard Chartered) have put Ant International's FalconTST 2.0 into live FX systems, moving a foundation model from benchmark to production.
2. The model forecasts cash flow and currency exposure, which feeds directly into hedging decisions and capital efficiency across a cross-border payments market projected to top $250 trillion by 2027.
3. Its reported MASE of 0.666 implies forecasts about a third more accurate than a naive baseline, but the benchmark result is based on Ant's own testing.
4. The technical advances (missing-data handling, cross-domain generalisation and multi-frequency support) target real weaknesses in financial time-series data.
5. The main risks are unverified performance claims, weakness at long horizons, and the systemic danger of many banks acting on correlated forecasts.

Frequently asked questions#

What is FalconTST 2.0 in plain terms? An AI model that studies histories of numbers moving through time and predicts what comes next: here, how much money will flow and in which currencies, so businesses can plan liquidity and hedge foreign-exchange risk.

How is it different from a chatbot model? Chatbots are trained on text. FalconTST is trained on numerical time series. Both use the transformer design, but one predicts words and the other predicts quantities over time.

What does a MASE of 0.666 mean? MASE compares a model's error to a naive forecast that repeats the last comparable value. Below 1 beats the naive rule; 0.666 suggests errors roughly a third smaller. It is scale-independent, so it works across different data types.

Is the benchmark claim independently verified? Not yet. Ant reports topping a leading public benchmark, but the result rests on its own testing. An API trial on GitHub allows outside developers to test the model.

Why should investors outside payments care? Forecasting quality affects hedging costs, liquidity buffers and capital efficiency across FX, money markets and treasury operations, and the deployment signals that foundation models are entering core bank infrastructure.

Could this create new systemic risk? Potentially. If many banks hedge from similar forecasts, they may take correlated positions that amplify market moves, a concern regulators already track under model-risk supervision.

Does this replace human traders? No. It produces forecasts that inform decisions; hedging and execution still sit with desks and their risk governance.

Glossary#

Time-series transformer (TST): A neural network of the transformer family applied to numerical data ordered through time rather than to text.

Foundation model: A large model pre-trained on broad data that can be applied to many tasks, including ones it was not specifically trained on.

Zero-shot forecasting: Predicting a new series without task-specific retraining, relying on patterns learned during pre-training.

MASE (Mean Absolute Scaled Error): A scale-independent accuracy metric comparing a forecast's error to a naive benchmark; below 1 means better than naive.

FX hedging: Using financial contracts to reduce the risk that currency movements erode the value of future cash flows.

Liquidity management: Ensuring an institution holds enough cash, in the right currencies and at the right times, to meet its obligations without holding excess idle funds.

GIFT-Eval: A public benchmark from Salesforce for evaluating general time-series forecasting models across many datasets and tasks.

CRPS (Continuous Ranked Probability Score): A metric that scores probabilistic forecasts, rewarding predictions that are both accurate and well-calibrated in their uncertainty.

References#

1. Ant International / Business Wire, "Ant International's FalconTST Model 2.0 Achieves SOTA…", 19 August 2026 (company announcement).
2. FinTech Futures, "Major banks tap Ant International's forecasting platform FalconTST 2.0", 25 August 2026.
3. TechNode Global, "Barclays, Citi, Deutsche Bank, Standard Chartered adopt Ant International's FalconTST AI model 2.0", 20 August 2026.
4. Bank for International Settlements, "Triennial Central Bank Survey of foreign exchange and OTC derivatives markets in 2022", 27 October 2022.
5. World Economic Forum, "Cross-border payments are estimated to be worth $250 trillion by 2027", 2023.
6. Aksoy et al., "GIFT-Eval: A Benchmark for General Time Series Forecasting Model Evaluation", arXiv (Preprint, not peer-reviewed).
7. Wikipedia, "Mean absolute scaled error" (summarising Hyndman & Koehler, 2006, International Journal of Forecasting).
8. Ant International, "Falcon-TST API trial", GitHub repository.
9. "Chronos-2: From Univariate to Universal Forecasting", arXiv (Preprint, not peer-reviewed).
10. "Moirai 2.0: When Less Is More for Time Series Forecasting", arXiv (Preprint, not peer-reviewed).
11. Monetary Authority of Singapore, official website (PathFin.ai programme reference).

This article is for information only. It is not investment, legal or financial advice, and nothing here is a recommendation to buy, sell or hold any asset. Performance and benchmark figures attributed to FalconTST 2.0 are claims made by Ant International and have not been independently verified.