Liquidity management was built around banking cycles that increasingly bear little resemblance to the way digital money moves. Customers transact around the clock, cross-border payments continue across time zones and digital banks can experience changes in cash flows long after traditional treasury windows have closed. Hong Kong regulators are now testing whether predictive AI can help close that gap, selecting Ant International’s Ant Bank and embedded finance provider Bettr for the liquidity risk management track of GenA.I. Sandbox++. The project will use time-series AI to forecast cash flows and foreign exchange requirements more precisely, with the longer-term objective of supporting treasury operations designed for a 24/7 financial system.
Ant Bank Brings Predictive AI Into Treasury
Ant Bank plans to deploy Ant International’s Falcon Time-Series Transformer 2.0 model to strengthen foreign exchange and liquidity management.
Rather than using generative AI primarily for customer-facing applications or employee productivity, the project applies AI to one of banking’s fundamental operational requirements: ensuring sufficient liquidity is available when and where it is needed.
Falcon TST 2.0 will generate granular daily forecasts intended to give Ant Bank’s treasury operation a more accurate view of future cash-flow requirements. Bettr’s Platform Tech will support deployment of the technology for the bank’s treasury team.
The companies aim to use the sandbox to establish practices for AI-driven liquidity risk management and ultimately support a financial environment in which treasury capabilities can respond to activity around the clock.
For Ant Bank, the requirement is particularly relevant because its digital banking model is closely integrated with an e-wallet ecosystem. Customer activity therefore does not necessarily conform to conventional banking hours, increasing the value of continuously understanding potential movements in liquidity.
Cash-Flow Forecasting Becomes an AI Use Case
Forecasting is already central to treasury management, but the growth of real-time and digital payments is increasing the amount and speed of information treasury teams need to interpret.
Unexpected changes in payment volumes, withdrawals or foreign exchange requirements can affect how efficiently a bank allocates liquidity. More accurate forecasts could potentially allow institutions to hold sufficient resources while reducing unnecessary buffers created by uncertainty.
Ant International originally developed Falcon for forecasting within real-world FX risk management across cross-border payments. Its latest Falcon TST 2.0 model is designed specifically for time-series forecasting rather than general-purpose language generation.
The company says the model achieved more than 93% forecast accuracy and state-of-the-art performance on the Mean Absolute Scaled Error metric on a public benchmark for time-series foundation models.
Those results are company-reported, and performance within a regulated bank treasury environment will ultimately be more important than benchmark results. The Hong Kong sandbox provides an opportunity to test how such forecasting capabilities perform when applied to actual financial operations.
24/7 Finance Changes the Treasury Problem
The project also highlights a less visible consequence of the shift towards real-time finance.
Much of the discussion around instant payments focuses on what customers experience: money arrives immediately, services remain available continuously and transactions no longer have to wait for conventional processing windows.
Behind that experience, banks still need to fund those transactions, manage currency exposures and maintain appropriate liquidity.
As payment systems increasingly move towards continuous availability, treasury infrastructure may need to become more responsive as well. A bank cannot offer genuinely real-time financial services indefinitely if its internal understanding of liquidity remains dependent on slower forecasting and decision cycles.
Predictive AI could help institutions move towards a more dynamic model in which expected cash flows and funding requirements are continually reassessed as new information becomes available.
That would make AI less of an interface sitting on top of banking and more of an operational capability embedded within the financial machinery underneath it.
Hong Kong Expands Its Regulatory AI Testing
GenA.I. Sandbox++ is being run by the Hong Kong Monetary Authority together with the Securities and Futures Commission, Insurance Authority, Mandatory Provident Fund Schemes Authority and Hong Kong Cyberport Management Company.
The initiative is designed to give financial institutions an environment in which AI applications can be developed and tested while regulatory, governance and operational considerations are examined alongside the technology.
Ant Bank and Bettr have been selected for the liquidity risk management track, giving the project a focus that differs from many of the customer-service and productivity applications that dominated the first phase of generative AI adoption in financial services.
The regulatory sandbox approach is particularly relevant for applications involving treasury and risk. Forecasting models may assist decision-making, but institutions still need controls around model performance, explainability, data quality, oversight and what happens when predictions prove incorrect.
As AI moves closer to functions affecting liquidity and financial risk, those governance questions become substantially more important.
AI Moves Deeper Into Banking Infrastructure
The significance of the Ant Bank project extends beyond whether one forecasting model improves the accuracy of a treasury team’s cash-flow projections.
Financial institutions have spent much of the current AI cycle experimenting with relatively accessible applications such as document summarisation, coding, customer service and employee assistants. The next phase is beginning to move AI into operational environments where decisions affect the actual functioning of the bank.
Liquidity management is one of those environments.
The potential benefits are significant. Better forecasting could improve capital efficiency, reduce unnecessary liquidity buffers and give treasury teams earlier visibility of emerging funding requirements. But the tolerance for error is also considerably lower than it is for an AI system drafting an email or summarising a document.
That changes what successful banking AI looks like. Model intelligence remains important, but reliability, governance and integration with existing risk controls become equally critical.
As digital banking and real-time payments push finance towards continuous operation, treasury may ultimately need to follow. AI-powered forecasting could become one of the technologies helping banks make that transition.
What it means for the industry
- AI is moving into core banking operations. Liquidity and FX management show how financial institutions are beginning to apply AI beyond customer service and employee productivity.
- 24/7 banking will require more responsive treasury infrastructure. Continuous payments increase the need for banks to understand cash-flow and funding requirements in near real time.
- Time-series models could become increasingly important in financial AI. Not every banking problem requires a general-purpose LLM, and specialised forecasting models may prove more suitable for treasury and risk applications.
- Better forecasting could improve liquidity efficiency. More accurate visibility of future cash flows may allow banks to balance resilience with the cost of maintaining excess liquidity.
- Governance requirements increase as AI moves closer to financial risk. Model accuracy, explainability, oversight and failure management become critical when AI informs treasury decisions.
- Regulatory sandboxes are becoming testing grounds for operational AI. Hong Kong’s approach could help establish practical standards for deploying AI in regulated, risk-sensitive banking functions.
Article Source: Alipay
