Kelvin Li Explains How Ant International Built FalconTST and Why It Gets Paid on Accuracy

1. Ant International built FalconTST 2.0, a specialised AI model for forecasting in finance, which has achieved state-of-the-art performance on the MASE metric. Take us inside that decision. What did Ant see, from where it sits across global finance, that made the specialist bet the obvious one?

Our big bet is that Large Language Models alone are not enough to solve problems in the financial industry. Hence, we developed a Large Data Model. LLMs predict the next number, LDMs predict the next data, which helps us to better manage our risks and price our financial solutions. It forms the core of financial decisions. Our view is that LDMs will form the backbone of AI for finance, and together with LLMs, this will form a holistic solution for global finance.

In fact, we recently unveiled our new AI-native end-to-end stack of financial solutions designed to serve global businesses. Our solutions span the full cycle of merchants' payment needs across payment, account, FX, treasury and growth. FalconTST forms the backbone of our Falcon Forecast treasury solution, which predicts cash flow, liquidity and FX positions for next hours to months, helping treasury teams plan ahead.

We developed FalconTST as we saw that there are still certain gaps in traditional banking services that cannot meet the needs of a digital economy – fixed banking hours, bank holidays. We decided to build a solution to predict our future cashflow so we can help our business and our customers better manage their global liquidity, foreign exchange, payments and other financial activities. This means we can settle faster with our customers and offer more competitive FX rates.

Today, FalconTST has over 8.5 billion parameters and has achieved over 93% accuracy in FX and cashflow forecasting, cutting corporate FX hedging and allocation costs by 30% to 60%.

2. Ant operates across payments, cross-border trade, treasury and digital banking on a global scale. FalconTST didn't emerge from a lab in isolation, it came out of a group that lives inside the real movement of money across all of those areas. What does building AI from inside an operating financial network let Ant see and do that a pure AI lab, or a bank, never could?

FalconTST can be called a “user-generated solution”, developed based on real pain points and real cases that we have seen from our businesses and our customers, refined by our financial data expertise.

Firstly, unlike the set-up of traditional banks, at Ant International our tech teams are integrated within our business and product teams. This structure allows us to work closely together to develop our in-house full stack of AI-native solutions.

Secondly, we have a different perspective from banks. We are building technology solutions for real-world impact for our business, and translating that into finance. We use overall financial intelligence, while banks adopt a more targeted approach grounded in their customers’ data. They are not as closely connected to the underlying commercial scenarios behind their customers’ transaction data, whereas FalconTST analyses the commercial activities underlying a business’s payments.

3. Four global banks now run FalconTST inside their own FX systems. What has Ant learned from being inside their operations that you didn't know before, about how the banking system actually forecasts, hedges and moves money at the frontier, that only a partner on the inside gets to see?

AI for finance is still in the early stages of development. From working with banks, we got a glimpse of how AI will interact with core banking systems – something we previously could not observe as we only interacted with supporting functions or operating functions.

4. From Ant's vantage point, where is financial AI actually going over the next few years? Not the general AI story, but the specific one you can see from inside global finance, what changes for banks, for businesses, for how money moves, once forecasting this accurate becomes normal rather than exceptional?

AI for finance will require decisive technology for a highly regulated industry, whereas current AI technologies such as LLMs are probability-based. One way that we solve the gap, is what we call a pay by performance model. For FalconTST, based on the model performance, we design a commercial mechanism. When the model accurately delivers a forecast result, we will be commercially rewarded. When it does not, we then take less revenue, or even no revenue. We align ourselves with the bank’s economic value, as the model’s forecast accuracy decides the bank’s risk and cost. If the forecast is accurate, their cost will be lower, and we get a higher share of revenue.

5. Ant has said more industry applications are coming beyond FX. Without asking you to give away the roadmap, what's the larger thing Ant is building toward here, and what does the financial system look like if that vision lands?

Firstly, we will continue to build on our model, following scaling law. We hope to be able to cover new industries through zero-shot use cases.

Secondly, we will also build a harness to allow FalconTST to adapt to different use cases, such as corporate treasury forecasting and FX forecasting.

6. Ant International sits at an unusual point in the system, a global fintech group that now builds the AI other financial institutions rely on. Looking five years out, what does Ant want to be to this industry, and what kind of financial system is Ant trying to help bring into being?

We see ourselves as the bridge between banks and businesses, working closely with regulators to support the scalable adoption of financial AI technologies. Our role is to leverage our financial expertise and technology and apply it to core areas of financial services. Banks can enhance their robust, secure solutions and businesses can enjoy more efficient, effective financial services at the same bank-grade security and compliance, which paves the way for responsible, industry-wide adoption of financial AI.