AI-driven financial services

The AI-driven Financial Service program strives to advance game-changing commercializable research in the spaces of retail lending, production & supply-chain financing, capital markets, institutional banking, and financial regulation. We support R&D efforts in broad areas of business and finance, emphasizing fresh approaches and novel solutions that are driven by AI and big data.

These include extracting business value from data that is transferrable across behavioural scenarios, developing augmented intelligence to facilitate credit and insurance decision-making, promoting novel practices of financial risk management to strengthen financial stability, and FinTech applications to hedging and financing in production and supply-chain.

From wealth management to blockchain and token technologies, we support cutting-edge research with commercial potential in optimization and machine learning algorithms, deploying reinforcement learning and deep neural networks.

AI-Enhanced Financial Technology

The AI-enhanced Financial Technology program commits to nurturing ground-breaking technologies with in-depth research potential that are relevant to contemporary business and finance.

We support projects with commercial potential in novel wealth management solutions utilizing advanced ideas from data science, optimization and control.

We deploy machine learning algorithms including reinforcement learning and deep neural networks, leveraging federated machine learning structures and their application in asset pricing and investor behaviour. We also support blockchain and token technologies for application in securitization and asset management; smart contracts, and digital currencies.

Social media provides a gold mine of data for understanding and forecasting financial market environments. We aim to leverage the power of AI-driven technology and the availability of social media big data to serve a wide range of clients from regulators to investors.


Social media provides a massive quantity of data, constantly updated, which may be used to understand and forecast financial market environments. This data directly affects how the economy operates, how investors make decisions, how regulators function, how the public responds, and how economies/markets interact. We aim to leverage the power of AI-driven technology and the availability of such data to develop and deploy a set of intelligence systems in three key domains; financial opinion, international financial relations, and the financial darknet.

These systems can be commercialized in two formats: syndicated products such as trend reports, and customized services such as live-monitoring visualizers.

P-1 AI-driven financial services

P-1.1: Managing Credit Risk for Online Consumer Lending

We take a data-driven approach to modelling consumer default risk in unsecured e-commerce lending scenarios. By drilling down to tick-level shopping behaviour in the financing records of a large population, we profile consumer credit at a granular level. Careful deciphering allows real-time assessment of future payment risk, particularly when payments are financed without posting collateral.

A typical cycle of credit-driven online shopping consists of three stages: browsing, placing an order and applying for payment credit, and paying instalments. Modelling challenges immediately arise because these high-dimensional records are massive samples exhibiting different degrees of serial-dependency at each event stage. They also show strong heterogeneity of mutual-dependencies across stages, and manifest a wide spectrum of event frequencies, ranging from browsing events hundreds of times a day, all the way to quarterly or semi-annual frequencies when periodical instalments are due.

One direction we plan to dive into is the construction of end-to-end deep learning models which fuse separate sequential neural networks, i.e. the Long-Short Term Memory model, to encode the microscopic action data separately for each of the three stages.

P-1.2: Forecasting Tail Risk in Financial Markets

This project investigates the forecasting problem of tail risk in financial markets and its implications for financial regulation. What keep risk managers awake at night are fast downfalls of unusual magnitude. These might trigger systemic spirals that could bring down the system. Dealers are concerned that market liquidity might dry up. And private investors sweat that the it will be difficult to climb out of the downside. Forecasting the far-left tail in any risk class or asset type is thus of great interest.

Predictions of near-extreme price movements require understanding of what drives tail dynamics. The static properties of return distribution are important and well researched in existing approaches, but that’s not enough. It is the inter-temporal dependencies that allow one to forecast. Although conditional quantiles oftentimes do move in sync with conditional volatility, tail dynamics may contain more than volatility clustering.

Across a wide range of assets, our preliminary studies show that the lengths of memory of conditional quantiles at different level indeed vary. Further, both far left and far right tail contain risk factors that are independent of those responsible for volatility clustering. This should have important implications for asset pricing in general and financial risk management in particular. We plan to substantiate this research by investigating various theoretical extensions and by applying the method to the real-world practice of financial risk management. Successful development of this project would result in novel business solutions that are readily useful to market participants as well as financial regulators.

P-1.3: Monetizing the Business Value of Consumer Behaviour

We examine consumer-profiling services offered by third-parties with a novel neural network structure and built-in interpretability useful to insurance companies for risk prediction and risk pricing. We test the effectiveness of the proposed methods on public datasets such as Kaggle and UCI, as well as on suitable real-world data from our industrial collaborators.

This project aims to develop methodologies and business solutions that address the problem of how to monetize behavioural data collected across different business scenarios, such as social media, e-commerce, etc., to support new business areas. One focus is how much business value can be added to insurance services using this approach.

The financial service industry is going through a paradigm shift from mass production to personalization. The immediate challenge facing service providers is that it asks for a much higher level of understanding of consumer behaviour to survive in a business environment that’s no longer product-centric. When financial services become data-driven, the services provided are naturally customized. It means a service provider must understand customer behaviour well enough before it knows who its customers are and how to identify them. One way to reach this level of understanding is through consumer-profiling services. And for the financial industry, machine-friendly labels interpretable by humans are essential when implementing financial models.

P-1.4: Analytics for Digital Relationship Management

We investigate the relative contributions of online and offline channels to the total value generated in financial services systems. This information can be used to evaluate and compensate the channels in a fair manner, and also to make better budget allocation decisions, helping firms to boost their revenues.

Many financial services firms still reach consumers through a combination of online and offline channels with the goal of acquiring new customers and managing relationships with the existing ones. Targeted online ads help businesses attract new customers whereas digital tools enhance the experience of existing customers. Traditional “human contact” methods are used when necessary to give a personal touch to the user experience. These various channels, when used in conjunction, may have a cooperative effect in creating a better customer experience.

To quantify the various contributions, we view the underlying channels as players in a cooperative game collaborating towards a common goal. We first propose a new metric that “adds” counterfactual reasoning to Shapley value while inheriting its desirable properties thereby addressing a known limitation of this cooperative game theory. Secondly, in order to make our proposal closer to reality, we show that a relatively broad class of cooperative games exists in which our proposed metric is amenable to efficient computation. We believe our theoretically-sound yet tractable approach has a unique edge over current practices and therefore, has the potential to be embraced by both researchers and practitioners. We also look at product-recommendation in settings as diverse as online retail, streaming services, and insurance and credit card recommendation.

P-1.5: Financial Decision Support for Managing Infrastructure Risks 

We offer a new decision-making framework for controlling the risk associated with insuring infrastructure projects for climate risk and technology obsolescence. We build tools that utilize insurance linked securities, and are based on the theory of real options.

Catastrophe bonds (also known as “cat bonds”) transfer a particular infrastructure risk, physical damage from natural disasters, from the insurer to the capital market. Climate risk changes over time due to anthropogenic factors, but also in an epochal or quasi-periodic regime-like manner over decades and across locations. This regime-like behaviour translates into fat-tailed, long-memory risk that may in some cases be predictable over the near-term.

The obsolescence of installed technology is another significant issue, and the cost sunk into the project cannot be recovered. Policy makers who seek to retire these assets to reduce carbon emissions, may face significant political opposition from owners, which may be publicly owned entities. In some ways these can be thought of as stranded assets, if free market principles were to apply.

The theory of real options may be a useful vehicle to consider the action set of financial and structural options for this class of problems, where uncertainties change and may resolve with time, for both shocks to the system and for the viability of technological options.

P-1.6: Counter-Cyclical Margin Calculation System for Option Portfolios

We develop a counter-cyclical margin calculation system app for option portfolios aimed at mitigating procyclicality whilst making full use of possible hedge properties within a portfolio. We also aim to present the supremum of the possible loss in terms of both the underlying asset’s future price and the time to maturity, and thus explicitly show the extent of the worst-case scenario. With this app, the regulator can observe the difference between the current market-condition-dependent margin level and the worst-case scenario loss, and adjust the emphasis on counter-cyclicality by modifying the parameters in our margin system.

Our approach differs from the most popular margin calculation currently used, which applies simulation methodology. This calculates the portfolio value among a finite number of selected scenarios and defines the margin of the portfolio as the maximal loss among these scenarios. While this risk-based approach is easy to implement, it depends on subjective scenario settings and thus market conditions. The approach is inevitably procyclical: the margin requirement is low in good times and high in bad times. As a result, traders receive margin calls and have to post additional collateral, often just at the time when a turbulent market makes it difficult for traders to raise cash or other liquid assets. To stabilize the financial markets, the financial industry has raised its concern on the procyclicality of such margin models. This project aims to answer these concerns.

P-1.7: Financial Sentiment Analysis and Market Prediction

We construct a real-time market sentiment index with three main components: sentiment extracted from social media, from market data, and option implied information. We investigate the interrelationships between these three components, generate an economically meaningful sentiment index for the whole market, and investigate its predictive power.

The market is driven not only by fundamental economic factors, but also by factors including social sentiment, political news and events. To better predict market movement, we need to study these multi-faceted influences. The most common strategy in current market prediction practice is to concatenate the features of various information sources into one super feature vector, which treats each information source separately and ignores their interactions. This practical challenge has led us to develop a supervised tensor regression learning approach in our investigation. Our tensor-based AI market predictor captures the relationships and interdependencies among various information sources. Furthermore, our AI market predictor adopts a sequence of tensors to reflect the time sequence of gathered information, as financial markets often experience sudden regime shifts near phase transitions due to financial crisis. In the face of a market turbulence, the correlation structure between stock prices changes. Our tensor-based AI market predictor enables us to empirically extract complex relationships from price time series and to predict abrupt changes by inferring the forthcoming dynamics of stock prices.

P-1.8: Hybrid Ensemble Modelling for Fraud Detection in Insurance Industry

This project develops a hybrid ensemble modelling approach that can help insurance companies combat fraud in more efficient ways. The key idea is to take advantage of both the business rules that are proven effective and the huge datasets that have now been accumulated in the insurance sector. Belief rule base (BRB) methodology is explored to develop a probabilistic rule-based model that can initially mimic and then fine-tune business rules established for fraud detection using labelled datasets. Pure data-driven machine learning (ML) models will also be developed, in order to detect fraud from different perspectives and to identify new fraud patterns from data. The two types of models: the BRB model and the ML model will be integrated by optimal training to develop a hybrid ensemble model and achieve the best accuracy and interpretability for fraud detection.

Ever since the introduction of insurance services in the mid-18th century fraud has existed, and it is now responsible for the loss of tens of billions of dollars each year. Increasingly, fraud detection services are seen to be a necessary component of a package of services offered by law firms. Identifying cases that should be disputed in the first place, is a key part of the process. Recent research has shown that computer algorithms can automatically search data and produce red flags that highlight cases where fraud is most likely. Most insurance companies already use automated red flags and business rules to assist in fraud prevention. Moving forward, this type of work can be done by more powerful AI decision systems, thereby making errors less likely.

P-2 AI-enhanced Financial Technology

P-2.1: Data-Driven Wealth Management

We apply robust optimization and machine learning techniques to develop data-driven asset management models that are free from parameter estimation and are relatively stable and insensitive to the random market environment.

Traditional wealth management is exclusive and expensive, and generally accessible to only high-net-worth individuals. Technology-enabled wealth management, exemplified by robo-advisors, significantly reduces costs and is both inclusive and scalable, thereby offering huge social benefits and business potential.

Most of the classical asset management models are variants of Merton’s expected utility maximization model and Markowitz’s mean-variance model, yet these models have major drawbacks. Firstly, their solutions are very sensitive to the key parameters, namely the mean and the covariance matrix of the stocks. Secondly, the true underlying probability is unknown in practice; so one has to resort to empirical versions of the mean and the covariance matrix. We project that our data-driven asset management models will provide a more accurate picture of the market environment.

P-2.2: Reinforcement Learning based Robo-Advising Apps under Enhanced Mean-Variance Framework

We develop a reinforcement learning robo-advising app based on an enhanced dynamic mean-variance framework. Robo-advisors offering digital wealth management services primarily targeted at retail investors represent a rapidly growing part of the wealth management industry. It is our explicit objective to ameliorate the shortcomings of time-inconsistency and non-monotonicity in current applications, while retaining the intuitive appeal, graphic representability of the probabilistic properties of the resulting wealth levels, and the elicitability of preferences by means of questionnaires. While predicting future return distributions represents the most compelling challenge in investment, any underlying distribution can be well approximated by utilizing a mixture distribution, particularly if we are able to ensure that the component list of a mixture distribution includes all possible distributions corresponding to the scenario analysis of potential market modes. Novel machine learning algorithms will be developed for this purpose.

P-2.3: Integrated Portfolio Management

We develop a portfolio management model based on an integrated analysis of asset return prediction, risk control and portfolio optimization. This project narrows the gap between theoretical research and practical applications, thus facilitating the commercialization of AI powered FinTech products. In practice there are three key factors for successful portfolio management: Asset return prediction, risk control and portfolio optimization. Practical portfolio management is in turn a combination of these three factors. Up to this stage, these three aspects have been typically investigated separately, thus causing shortcomings in each traditional model. We aim to supersede this methodology with an integrated model.

Traditional forecasting methods are mainly econometric models based on historical data, but these models are actually “backward-looking” methods. Investment, on the other hand, is a “forward-looking” activity, and the market is relatively effective, which means that the “backward-looking” prediction is invalid, at least theoretically. In terms of risk control, in traditional portfolio models there is usually only single total risk measure, which is unable to cope with the requirements of the actual portfolio risk management. In terms of portfolio optimization, most models simply consider “optimization” itself. In practice, portfolio optimization should combine prediction and optimization with respect to asset class and risk measure.

P-2.4: Likelihood Inference and Interpretable Machine Learning for Risk Analysis and Loan Decision Making

We develop an integrated interpretable machine learning (ML) model and AI system to support efficient, consistent and informative loan decision-making. Loan decision-making, also known as underwriting, is a key process of operations management in commercial banks. New technology and ML algorithms have opened the door for more efficient loan application and approval process compared to the traditional manual process. Digital technology advancements allow the use of big data to find multitudinous ways in which relevant factors influence each other. Complex relationships among many factors can be found by advanced ML models, which can aid more accurate classification of the risk status of loan applications.

While AI systems have huge potential to revolutionise loan decision-making, it is increasingly recognised that to be successful, it is vital to explain decisions made by ML models and their inner workings. We address this concern by developing an integrated interpretable ML model and AI system to support efficient, consistent and informative loan decision-making. Factors in the ML model are selected, measured and structured on the basis of underwriter guidelines and default rules, while nonlinear relationships among the factors are explored by data analysis and likelihood inference. A unique feature of this modelling strategy is that the integrated ML model makes full use of data analysis to capture nonlinearity among relevant factors. The model can also be optimally trained to achieve high accuracy in making correct loan decisions, whilst the loan decision-making process remains highly interpretable according to the guidelines and rules embedded in the model.

P-2.5: Development of DeFi Platform, Stablecoin, Option Smart Contract and Other Digitized Products

Decentralized Finance (DeFi) is increasingly used to create financial derivatives on blockchain, but the lack of stable digital basis assets currently implies considerable risk. We propose a stable coin mechanism as the fundament of the DeFi system. The platform can issue stable coins through matchmaking tradeoff.

An option smart contract, a typical financial derivative based on the stablecoin, gives its owner the right to receive a specified amount of underlying assets in exchange for basis tokens during the period before expiration at a specified price (strike price). The transaction that uses this right is called an “exercise.” An anti-option contract gives its owner the right to receive the “non-exercised” part of the deposited underlying and the “exercised” part of basis token after the expiration date.

The significance of this decentralized and programmatic derivative lies in the underlying asset not being held by any entity or intermediary. Instead, it is governed by scripts running on blockchain and will only be executed according to predetermined business logic. This will mitigate third-party risk and guarantee transparency. These problems were never fully solved by the financial markets until blockchain technology came along. Countless opportunities exist with this product, and many other derivatives and financial tools can be added to the ecosystem once we have built up the core capability.

This project has strong commercial prospects as digital currency and digital assets have been gaining increasing popularity as part of asset portfolios in recent years. DeFi is deemed to be the future financial model and the platform is best realized through blockchain technology. A well-designed stablecoin and trading mechanism will fuel the platform and ensure its robustness. Various financial products can thenceforth be “spun off.”

P-2.6: An Enterprise-grade Solution for Digital Asset Safeguarding and Transaction

We propose an enterprise-grade solution for digital asset safeguarding and transaction through the integration of end-to-end hardware and software security, thus achieving scalability, availability, security and governance of crypto funds in one package.
In the past few years, more than USD 1.2 billion in digital assets have been lost due to insecure usage of tools for holding private keys. A private key is a unique series of hexadecimal strings derived from asymmetric encryption, and it is used to make a digital signature when initiating a transaction. It is widely recognized that whether the owners of cryptocurrencies store their private keys in an isolated environment or in a place that is network-accessible, it is unlikely that flexibility for daily use will be maintained without compromising security.

Private keys are sealed offline in the Hardware Security Module (HSM), and multi-firewalls are established to provide advanced protection against common threats, including physical risks, internal risks, third-party risks and phishing attacks. The initialization and restoration of configuration, as well as the signing of a transaction, are fully completed in an isolated space offline, thereby providing maximum protection against software virus. Duty segregation is carried out by multi-party authorization, while the governance rules are kept in the form of scripts to enable a customized governance policy in assigning roles, limit rate, and a time lock. A transaction initiated by any party must be approved according to the governance before it is signed and broadcast to the blockchain.

To improve its usability and feasibility in common business scenarios, a temperature modifier is provided to bridge the gap between the hot and the cold wallet and allow parts of the fund to circulate without manual procedures and hardware authentication. In any case, the shared owners can revive the private keys of the enterprise digital wallet by combining their individual private keys, while the operators must perform as the predetermined administrative rules request. This strategy of duty segregation offers clear responsibilities for participants involved in crypto-asset management and ensures an effective and secure custody solution for enterprise use.

P-2.7: A blockchain-based data sharing platform for financial applications

This project aims to establish a secure and verifiable data sharing infrastructure with built-in security and privacy using blockchain technology. By storing and sharing encrypted digests of financial data on a blockchain platform, authorized institutions can properly trace historical transactions without relying on the involvement of third-party auditing. Our blockchain-based data sharing platform can break data silos in the present-day financial sector and boost the efficiency of a wide swath of services.

The main challenge of designing a blockchain-based data sharing platform is providing strong protection mechanisms to guarantee on-chain data confidentiality. Although applying standard encryption schemes can protect data privacy, it has become a challenging issue for users to search and share conveniently and efficiently over encrypted data. Meanwhile, as financial data usually contains rich content, it is often necessary to have more expressive searching algorithms. This is the dilemma between data confidentiality and data utility.

Another challenging issue for this project lies in the problem of query authorization relating to policy compliance. In practical deployment, query authorization refers only to authorized users who have access to data records. For example, in a likely scenario in credit assessment, authorized decision-makers’ access is limited to search within the applicant’s previous financial documents concerned with credit transactions. Query authorization needs to urgently address this issue.
We tackle these challenges from two directions: Firstly, for secure and efficient data sharing over the blockchain platform, we investigate a hybrid storage architecture where raw data is stored off-chain on dedicated storage servers and a secure index is kept on-chain for encrypted search. Secondly, for query authorization, we plan to design a fine-grained access control layer on top to safeguard data sharing with policy compliance.

P-2.8: AI Solver for Hard Financial Decision Making

This project aims to develop efficient AI solution methodologies for solving long-standing financial optimization problems and challenging issues that are emerging. The final deliverable will be a software system which includes a family of solution packages for different financial decision problems.

The past half-century has witnessed a remarkable advancement in modern financial decision theory and practice, both in enhancing our fundamental understanding of market randomness, and in enabling market participants to harness appropriate risk for a better return.

Still, many financial decision-making and risk management problems remain open and challenging, and are nonconvex in nature. Essentially, if we incorporate real world considerations into our financial decision-making models (for example, investors’ asymmetric risk attitude towards gain and loss), we will be surrounded by a nonconvex world. One example is the mean-Value-at-Risk (VaR) portfolio selection problem, which is both nonconvex and discontinuous. In fact, the real financial world generates endless lists of nonconvex optimization problems, including portfolio selection under high-moment risk measures (skewness and kurtosis for example), cardinality constrained portfolio selection, risk parity strategy, tax-loss harvesting, and fixed income portfolio.

A long list of efforts from optimization, statistics and machine learning has led to relatively efficient algorithms for solving nonconvex optimization problems: stochastic gradient descent, momentum regularization, variance reduction, and mini-batch optimization. Also neurodynamics-based portfolio optimization deserves in-depth investigation in its own right because of the distinctive complexities in depth and scale of financial engineering and management.

P-3 Social Media Analytics for Financial Technology and Services

The Social Media Analytics for Financial Technology and Services program launches with the following 3 projects:

P-3.1: Detecting and Predicting Public Opinion on Financial Environment

P-3.2: Detecting and Predicting International Financial Relations

P-3.3: Detecting and Predicting Financial Darknet Activities

P-3.1: Detecting and Predicting Public Opinion on Financial Environment

Social media has proven to be a highly cost-effective alternative to traditional monitored mechanisms for detecting and predicting public opinion trends in financial market environments. In this project, we draw on our long-standing research experience in collaboration with DataStory (a social media market research lab) to develop initiatives in two directions. Firstly, to build a data collection capability for real-time public opinion data from news media, public forums, blogs, social networks, and other user-generated content platforms. Secondly, to use computational social science methods (combining social science theory of public opinion and machine learning algorithms) to mine the collected opinion data and derive predictive models of financial public opinion dynamics. Finally, we will produce and commercialize a financial public opinion monitoring system based on the resulting predictive models coupled with constantly updated social media data.

This financial public opinion monitoring system will potentially serve a variety of institutional clients, including investment institutions, consulting firms/think tanks, policy makers, academic research units, financial media firms, NGOs, etc. The system can be commercialized in two formats: syndicated products such as daily or weekly trends reports on a subscription-fee basis, and customized services such as live-monitoring visualizers tailored made for specific clients. We will launch the products and services first in Hong Kong, and gradually expand to other parts of the world.

P-3.2: Detecting and Predicting International Financial Relations

AI technology has proven to be an effective and efficient tool to monitor international financial relations for countries, organizations, and individuals alike. In this project, we plan to build a real-time data collection capability for data on international financial events as well as issues from news media, world organization sources, and various financial social media and user-generated content platforms. We will then use computational social science methods (combining social science theory of international financial relations and machine learning/knowledge graph algorithms) to mine the data and derive predictive models of international events/issues. This will enable us to build a knowledge graph of world financial events and provide a baseline infrastructure for international financial relations monitoring. We will then produce and commercialize an international financial relations monitoring system based on the predictive models, coupled with a constantly updated international financial relations data.

This AI-enabled and knowledge-graph supported monitoring system of international financial relations will potentially serve a variety of institutional clients, including, investment firms, consulting firms, think tanks, academic research units, financial media, NGOs, etc. The system will be commercialized in two formats: syndicated products such as daily or weekly trends reports on a subscription-fee basis, and customized services such as live-monitoring visualizers tailored made for specific clients. We will launch the products and services first in Hong Kong and gradually expand to other parts of the world.

P-3.3: Detecting and Predicting Financial Darknet Activities

A wide range of negative activities such as misinformation, spamming messages, hack attacks, online frauds, etc. can originate from financial darknet services which are typically paid for in cryptocurrencies. Although the technical community has made efforts to understand how some of the financial darknet activists operate, there is a lack of comprehensive framework for monitoring financial darknet contents and associating them to cryptocurrency transactions for detecting potential negative activities.

In this project, we will build a data collection capability on dubious/negative activities from both open social media and hidden darknet such as TOR networks. We will then use computational social science methods, combining social science theory of deviation behaviour and machine learning/knowledge graph algorithms, to mine the collected darknet data and derive predictive models of darknet events/issues.

We will then build a real-time data collection capability on cryptocurrency transactions and use machine learning algorithms to detect suspicious transactions and associate them with darknet events/issues. Finally, we will produce and commercialize an interactive risk activity monitoring system, based on the resulting predictive models coupled with updated financial darknet data and cryptocurrency transactions.

This financial darknet and cryptocurrency monitoring system will potentially serve a variety of institutional clients, including regulators, investment firms, consulting firms, think tanks, academic research units, news media, NGOs, etc. The system can be commercialized in two formats: syndicated products such as daily or weekly trends reports on a subscription-fee basis, and customized services such as live-monitoring visualizers tailored made for specific clients. We will launch the products and services in Hong Kong and gradually expand to other parts of the world.