Ant International’s FalconTST 2.0 Advances Predictive AI for Finance and Global Business
Ant International has unveiled Falcon Time-Series Transformer (FalconTST) Model 2.0, the company’s latest artificial intelligence model designed to improve forecasting accuracy for complex, rapidly changing time-series data. The new model is initially focused on practical applications in foreign exchange (FX) risk management and cross-border payments, while Ant International plans to expand its use into areas such as e-commerce demand forecasting, supply-chain management and aviation operations.
The launch represents another step in the development of predictive AI for financial services. While generative AI and large language models have transformed how businesses process and generate text, FalconTST is designed around a different challenge: understanding numerical information that changes continuously over time and using those patterns to anticipate what may happen next.
For financial institutions and global businesses, that capability can have a direct impact on liquidity planning, currency management, capital allocation and operational efficiency.
FalconTST 2.0 Reaches State-of-the-Art Performance
One of the key highlights of FalconTST 2.0 is its performance on the Mean Absolute Scaled Error (MASE) metric, a widely used measurement for evaluating the accuracy of time-series forecasting models.
Ant International said FalconTST 2.0 achieved a MASE score of 0.666 on a leading global public benchmark for time-series foundation models. The result places the model at the top of the benchmark leaderboard and ahead of other time-series transformer models developed by major global technology companies.
MASE is particularly useful when comparing forecasting performance across different datasets because it measures prediction errors relative to a baseline. A lower MASE score generally indicates more accurate forecasting.
For Ant International, the benchmark result is significant because the company’s objective extends beyond achieving strong performance in an academic or testing environment. The company is seeking to apply time-series AI to real-world financial challenges where even relatively small improvements in forecasting can influence major decisions involving liquidity, foreign exchange and transaction management.
AI Designed for Numerical Financial Data
Financial systems generate enormous quantities of numerical data every day. Transaction amounts, account balances, payment flows, settlement requirements, currency positions and liquidity needs can all change continuously.
This type of information presents a different challenge from the text-based information used by large language models. Rather than understanding words and their relationships, time-series models need to identify patterns that develop over minutes, hours, days, weeks or months.
In global payments, accurate forecasting can help organizations determine when funds will be required, how much liquidity should be available and which currencies will be needed.
This is especially important for businesses operating across multiple countries. A company may receive payments in several currencies while simultaneously making payments, settling invoices or covering expenses in other currencies. The mismatch between incoming and outgoing currencies can create foreign exchange exposure.
Forecasting therefore becomes an important component of FX risk management.
For example, an airline may receive ticket revenues in U.S. dollars, euros, pounds and other currencies while its expenses include aircraft leasing costs, airport charges, fuel expenses and supplier payments denominated in different currencies. To manage this exposure, the airline may use FX hedging strategies.
However, effective hedging depends partly on understanding future cash flows. If the business overestimates the amount of foreign currency it will receive, it could hedge more than necessary. If it underestimates future requirements, it could remain more exposed to unfavorable currency movements.
FalconTST is designed to improve this forecasting process by identifying patterns in historical and current data and using them to estimate future cash flows.
Moving Beyond Traditional Forecasting Models
Historically, organizations have often developed separate forecasting models for individual business functions.
A retailer might build one model to predict product sales. An airline could use another system to estimate passenger demand. A bank may have separate models for liquidity forecasting, transaction volumes and cash management.
These models can be effective for specific applications, but they can also require substantial customization and retraining when businesses enter new markets or face new forecasting requirements.
Time-series foundation models take a broader approach.
Rather than being trained exclusively for one narrowly defined task, FalconTST is designed to learn general patterns that occur across different types of time-series data. These patterns can include trends, seasonality, recurring cycles, changes in demand and sudden shifts in activity.
Although financial transactions, airline bookings, retail sales and energy consumption may appear unrelated, their underlying time-dependent behavior can contain similar structures.
By learning those common characteristics, a foundation model can potentially be adapted to new forecasting scenarios without requiring an entirely new model for every application.
Ant International says this broader capability is central to its strategy for FalconTST.
From Internal Financial Management to Global Banking Applications
FalconTST was initially developed and deployed within Ant International to support its own cash-flow and FX exposure management.
The model is used across different forecasting horizons, including hourly, daily and weekly requirements. This allows Ant International to monitor and anticipate changing liquidity and foreign exchange needs associated with its global cross-border payment activities.
The technology has also been integrated into FX and liquidity management solutions involving major international banking institutions, including Barclays, Citi, Deutsche Bank and Standard Chartered.
Barclays has integrated FalconTST into its BARX NetFX FX hedging platform, while Citi combines the model with its Fixed FX Rates solution. Standard Chartered uses FalconTST alongside its SCALE FX system as part of its participation with Ant International in the PathFin.ai program supported by the Monetary Authority of Singapore.
These applications demonstrate how predictive AI can move from an experimental technology into operational financial infrastructure.
According to Ant International, these banking applications have adopted FalconTST 2.0, with the company reporting forecasting accuracy improvements of more than 93% consistently.
For organizations processing large volumes of international transactions, better forecasting can translate into more effective liquidity planning and improved management of foreign exchange exposure.
Addressing the Challenges of Real-World Data
One of the most important aspects of FalconTST 2.0 is its focus on practical data challenges.
Real-world financial datasets are rarely perfect. Data can be missing because of reporting gaps, system interruptions, weekends, holidays or differences in transaction schedules. A forecasting model that interprets every missing observation as a zero could generate inaccurate conclusions.
FalconTST 2.0 is designed to distinguish missing observations from genuine zero values.
This distinction can be particularly important in financial forecasting. For instance, a lack of recorded transactions during a weekend should not necessarily be interpreted as an indication that future demand will remain at zero. Understanding the reason behind an absence of data can prevent the model from learning misleading patterns.
The updated model also focuses on generalization across different business domains.
Through its ORBIT approach, FalconTST is designed to learn common time-series characteristics from areas including finance, retail, energy and tourism. This enables the model to apply its forecasting capabilities to new scenarios rather than remaining restricted to the environment in which it was originally trained.
Supporting Multiple Time Horizons
Another technical feature of FalconTST 2.0 is its ability to work with multiple time frequencies.
Different industries require forecasts at very different intervals.
Payment systems may need to analyze transaction data at the second level. Treasury teams may require hourly forecasts to manage liquidity. Airlines may need daily demand forecasts to prepare for passenger volumes. Economic and business planning models may operate on a monthly basis.
A forecasting architecture capable of handling multiple frequencies can therefore support a wider range of applications without requiring organizations to maintain completely separate systems for every time horizon.
This flexibility is particularly relevant for global businesses, where operational requirements can change dramatically depending on the type of activity being forecast.
Expanding Into E-Commerce and Aviation
Although financial services and payments remain central to FalconTST, Ant International sees opportunities to extend the model into other industries.
E-commerce is one potential application. Online platforms need to predict customer demand, transaction volumes, inventory requirements and supply-chain activity. Demand can change rapidly because of promotions, holidays, market conditions and consumer behavior.
More accurate forecasting can help businesses determine how much inventory they need and when additional resources should be allocated.
Aviation presents another significant use case. Airlines operate highly complex international businesses in which revenues, expenses and operational requirements can fluctuate across multiple currencies and geographic markets.
FalconTST can potentially support forecasting across passenger demand, cash flows, FX exposure and other operational requirements.
The same principles could eventually be applied to logistics and other industries where businesses need to make decisions based on constantly changing numerical information.
Turning Predictive Intelligence Into Business Decisions
Ant International’s broader vision for FalconTST is not simply to produce better forecasting numbers. The company aims to turn predictive intelligence into actionable business decisions.
Jiang-Ming Yang, Chief Innovation Officer at Ant International, said that while large language models have demonstrated AI’s ability to understand and generate information, FalconTST focuses on another important capability: understanding how conditions evolve over time and anticipating future developments.
The practical value of this capability lies in decisions such as how much liquidity a company should prepare, how much FX exposure it should hedge and how capital should be allocated.
Kelvin Li, General Manager of Platform Tech and Senior Vice President at Ant International, similarly highlighted the operational benefits of forecasting for global businesses. According to Li, FalconTST has already helped clients improve forecasting and achieve cost savings, while the second-generation model is intended to extend those benefits to banking partners and businesses operating in fast-moving sectors such as e-commerce, travel and fintech.
From a Financial Model to a Broader AI Foundation
FalconTST 2.0 reflects a wider shift in enterprise AI toward specialized foundation models designed for specific types of data.
Large language models have become increasingly important for text, images and other forms of unstructured information. Time-series foundation models address a different category: numerical information that changes over time.
For financial institutions, this distinction is particularly important. Markets, payments, liquidity and currency exposures are dynamic. A forecast that is accurate today may need to be recalculated tomorrow as transaction patterns and market conditions change.
The ability to continuously interpret those changes and provide forecasts could make predictive AI an increasingly important component of financial infrastructure.
Ant International’s strategy is also based on making FalconTST reusable. Instead of creating a separate forecasting system for every customer or industry, the company wants to develop a common predictive capability that can be adapted to different business environments.
That could allow the same underlying technology to support FX management for a payment institution, demand forecasting for an e-commerce platform or operational planning for an airline.
The introduction of FalconTST 2.0 highlights how AI development is expanding beyond generative applications toward systems that directly support business forecasting and decision-making.
Its reported benchmark performance, combined with applications in FX risk management and liquidity forecasting, positions FalconTST as an example of how specialized AI models can address complex enterprise requirements.
The next phase will likely depend on how effectively such models perform across increasingly diverse datasets and real-world operating environments.
For Ant International, FalconTST 2.0 is intended to serve as more than an improvement over its predecessor. The company is positioning the technology as a reusable predictive AI capability that can support financial institutions and businesses across multiple sectors.
As cross-border commerce continues to expand and companies manage increasingly complex international cash flows, the ability to forecast financial and operational needs accurately could become a significant competitive advantage.
FalconTST 2.0 represents Ant International’s effort to build that capability, beginning with payments, FX and liquidity management and potentially extending into e-commerce, aviation, logistics and other data-intensive industries. Its development illustrates an emerging direction in enterprise AI: not simply generating information, but helping businesses anticipate what comes next and make more informed decisions before those changes occur.
Source link: https://www.businesswire.com
