When the United Kingdom cast its decisive vote on 23rd June 2016 to leave the European Union,

a membership in which it held for

more than 40 years, the British pound slumped to a 31-year low as the final polling results sent shockwaves during the Asian trading
hours. The losses extended to the European and US trading sessions as panicking investors fled to safe haven assets, and stunned
traders caught short by the unexpected outcome rushed to cover their positions. On that day, the pound plummeted more than 10%
to $1.33, from $1.50.

While the financial markets absorbed the news and braced for further turmoil over the following days and weeks, no one was quite
prepared for the “flash crash” that happened 3 months later, on 7th October, when the currency plunged within a few minutes from
$1.26 to $1.15 – marking a fresh 31-year low.

The blame swiftly shifted to “algorithm trading programs”, for triggering market orders that contributed to the massive pressure on
the pound as political uncertainties mount.

Algorithm-driven robot traders

Algorithm-driven robot traders, a form of “Artificial Intelligence (AI)”, mimic real-life trading using logic, if-then rules, decision trees to
behave in ways that resemble an expert trader.

Initially developed to improve trading efficiency by minimizing the manual tracking of financial markets and laborious execution of
order (and arguably, also to eliminate trader emotional volatility), these robo-trading algorithms have evolved. From simple sell-buy
triggers, to devising trading strategies built on high-speed cross-asset-correlations and other complex mathematical calculations, they
have acquired the potential to create systemically contagious impacts as trades from one algorithm could trigger signals of others (as
we see in this Brexit example).

The coding of the financial markets data tracking and profitable trades structuring is not new; what’s changed is that these algorithms
fully harnessed the vast computation power available today to rapidly identify micro arbitrage opportunities across assets, markets,
time zones and construct profitable trading strategies within fraction of a second.

Processing power, and lots of data

“Artificial intelligence” encompasses a vast range of technologies, ranging from problem-solving programs that copy human logical
thinking process (as in this case Algorithm-driven robot traders), to “machine learning” that improves these programs over time (“with
experience”) using mathematical optimization techniques, to “deep learning” (or deep neural networks as formally referred to in
academic research) which are composed of multi-layered neural networks that self-train with vast amounts of data. In the fields of
speech and image recognition, for example, Amazon’s Alexa, Apple’s Siri, Microsoft’s Cortana, and the many voice-responsive features
of Google – are enabled by the vast computation power as well as volumes of image, video, audio and text file data available on the
Internet.

There is no question that it is in the machine-vs-human game of chess where this impressive processing power has taken our
appreciation of potential of AI to the next level. Deep Blue (IBM’s supercomputer) beat Garry Kasparov, the then world chess
champion, in a six-game match in 1997, by using sheer processing power and massive data storage capability. Moving beyond merely
programming how human experts think with if-then-rules and decision trees, Google’s AlphaGo (an application of two layers of deep
learning nets – Deepmind combined with a reinforcement learning) played against Mr Lee Se-dol last year in the ancient Chinese game
of GO. AlphaGo beat Mr Lee, perhaps the best player of the game, in four of the five games.

These advances in AI are made possible by the increased computational power referred to as Moore’s Law and graphics processing
units (GPUs) – initially built by Nvidiá for 3D visual experiences in gaming - which enable 20 to 50 times efficiency compared to
traditional central processing units (CPUs). Google’s tensor processing units (TPUs), or Intel’s acquisition of Nervana Systems and
Movidius, two startups that tailor-make technology for deep-learning computations point to how serious technology giants are viewing
the potential in this market.

Sheer processing power combined with the availability of realms of data are accelerating AI applications across industries. Besides
robo-trading, we are seeing innovations in the areas of robo-
advising, fraud detection and market behavioral analytics in the
financial services.

Spot Light
Artificial Intelligence in the Financial Services
OCBC Bank launched the OCBC OneWealth mobile app in March
Robo-Advisors offer digital investment advisory services based on 2016, an investment one-stop-shop that provides customers with
algorithms. By collecting the details of investors’ investment convenient access to market information, investment ideas,
objectives, preferences, style and risk profile, the robo-advisers personalised alerts on their existing investments, and even the
learn what investors are interested in and deliver customised advice ability to directly purchase unit trusts on their mobile devices.
by aggregating relevant research reports and market updates to Customers can login conveniently to the app with just the touch
of their fingerprint (via OCBC OneTouch biometric authentication
suggest financial asset allocations.
technology) to see all their investment holdings and share articles
from the OCBC OneWealth app on social media apps including
In addition to these data analytics approaches, robo-advising
Whatsapp, Facebook, LinkedIn and Twitter. Customers who are
technologies such as Chatbots (robots that converse with humans)
new to investing, or who are exploring investment ideas, can
or Sentiment Analysis (the “irrational and qualitative” aspect of
access the app's curated content, information on the latest
investment analytics, based on non-balance-sheet components such macroeconomic trends and analysis by OCBC Bank experts to
as views sourced from Tweets or other social media) which improve better understand market situations through expert analysis with
the customer experience with natural language processing and up-to-date market insights every morning. This provides
unstructured data analytics algorithms, have also being widely customers with a simple and convenient way to manage their
deployed. wealth and investments. The OCBC OneWealth app is also the
only mobile channel in which customers can purchase and
This robo-human interaction technology is in initial phases of redeem Singapore Savings Bonds, giving them the flexibility to
innovation. Robo-adviors are yet to understand subtleties in a bank on their phones. This “democratises” wealth management
conversation. “I am worried about my parents’ health” which may and makes access to investment information and wealth products
prompt a human advisor to review the risk profile and investment easier and within reach of more customers.
horizon of the customer, may not necessarily trigger the same
Saxo Bank - Addressing the large pool of self-directed traders,
response in a robo-advisor. A robo-advisor may also be limited in its
early accelerators of Fintech are coming to the table with
information gathering ability: it may not ask about money held solutions addressing previously time-consuming and cost-
outside of its service, which could give a distorted picture of a intensive middle-man challenges. Recognizing that in an era when
customer’s financial health. artificial intelligence drives high frequency trading, Saxo Bank,
parent company of Saxo Capital Markets, has built a one-of-its-
These examples show that whilst there is still some way to go kind comprehensive price sourcing engine to indicate the current
before a robo-advisor can fully function as fiduciary in the best price for an instant Bond transaction, based on 40 global
traditional sense, the volume and speed of the data being processed bonds providers and major exchanges called by its Bond Robotics
algorithm. This represents a marriage of state of the art tech and
across several sources to deliver timely advice mean that
existing institutional tools to democratize investment for the end
innovations in these technologies will continue. Certainly, for those user client.
contemplating using robo-advisers, less biased advice combined
with a wider selection of potential investments at a fraction of the
cost of traditional service is an attractive proposition.

Artificial Intelligence in the Financial Services

Algorithmic Trading by Robo-advisors are a class Fraud Detection Analytics include Market Behavioral
computers which are programmed to of financial adviser that provide Supervised Learning with known characteristics,
take certain actions in response to financial advice or portfolio where classification or regression based approaches
Analytics monitors financial transactions
varying market data: ‘algorithmic management with minimal human are used to build fraud patterns; or Unsupervised for earning warning signs such as cancel-and-
trading relies on computer systems to intervention, using algorithms to Learning, where in the absence of pre-defined amend trades or aberrations of trading
buy shares automatically when automatically allocate, manage and known characteristics, learning of suspicious patters patterns to detect unauthorised and rogue
predefined market conditions are met’ is via clustering, exploratory, or data reduction trading situations.
optimize clients’ assets.
-- Oxford Living Dictionary approaches.
Fraud Detection - AI machine learning techniques are also used to
help in fighting cyber attacks, through automatic scanning, detection
Spot light
and response of network vulnerabilities. Similarly, by applying AI to
volumes of data to spot suspicious financial transactions amongst Westpac Banking Corp together with National Australia Bank and
millions of normal ones, AI could ease the burden on investigators in Qantas – invested in the start-up Data Republic last year with the aim
combatting money laundering, financial fraud and sanctions to allow companies to securely exchange data with each other to
violations. enhance customer offerings.

With increasing regulatory scrutiny in these areas, financial One potential application of data exchange is to deliver dynamic
institutions have adopted over-cautious attitudes, setting thresholds services tied to customer’s travel schedules. For example, from
of traditional rules-based anti-fraud systems at levels that raise alert Qantas’s data, Westpac would be able to correlate customer’s
on practically everything resulting in unsustainable increase in false movements to their potential funding requirements - for example,
ensuring that their credit cards are not switched off for suspected
positives. Not only do legitimate customers face unnecessary probes,
fraud as transactions start appearing in that foreign country.
investigators also consume excessive time clearing these false
positives. Adding to this workload is the manual building of the
customer profile when swamped with structured and unstructured
data about the subject, their social and commercial networks from in-house and other public and commercial sources.

By replicating the way an investigator manages a case, AI automatically flags unusual/suspicious activity by mining data from a
customer’s and peer group transaction history and thousands of “signature fraud patterns”. At the same time AI also learns new
patterns or goes into corrective loop to ignore the ‘false positives’. For investigators facing the tedious job of manual data collation
and rules update in the legacy threshold systems, AI not only reduces the burdens but also completes these tasks much quicker.

Market Behavioral Analytics - In the fast-paced, high-pressure

world of trading where it is not uncommon for millions of
transactions to change hands across the global markets of FX, Spot light
futures, or commodities, most would rank Nick Leeson and the
collapse of Barings Bank, the United Kingdom's oldest merchant bank Signac, a joint venture between Credit Suisse and Palantir
in 1995, as one of the most publicized cases of unauthorized trading. Technologies, the Artificial Intelligence firm backed by the CIA’s
Venture Capital arm, uses data-driven behavioural analysis to root
Trading in the futures markets on the Singapore International out rogue traders.
Monetary Exchange (SIMEX), Leeson was regularly using Barings'
error account (accounts used to correct mistakes made in trading) Wall Street has been anxious to improve surveillance after a string
numbered 88888 to hide his trading losses, a practice that remained of unauthorized trading scandals. The UBS and SocGen trading
undetected for at least 2 years. The unravelling was triggered by his scandals shared factors, which Signac called “toxic combinations”,
including trading in internal accounts and a large number of
attempts to offset losses when the 17 January 1995 Kobe earthquake
“cancel and correct” trades, could be used as early warning signs.
struck sending the Asian markets and his trading positions into a
tailspin. His new trades exacerbated the original losses, the total of Signac works with financial institutions to mitigate and prevent
which eventually reached £827 million (US$1.4 billion), resulting in unauthorized activity by employees, like rogue trading and other
Barings declaring insolvency on 26 February 1995. fraud. Signac integrates disparate data from disconnected, legacy
IT systems, so different datasets like HR data and profit and loss
Recent cases of unauthorized trading included Jérôme Kerviel, a data can be examined in one place to look for irregular patterns.
French trader convicted in the 2008 Société Générale €4.9 billion Signac compares employee behavior to their historical activity and
trading loss scandal. As a trader at the bank's Delta One desk, he that of their peers. They then surface data indicative of suspicious
created offsetting faked hedge trades to cover his losses. Three years behavior, proactively notify those responsible, and prompt them
later in 2011, in what was another incident of unathorised trading to review the alerts and take action. Signac's Investigative Unit also
loss, Kweku Adoboli, as a Global Synthetic Equities desk trader at uses Signac technology to review and test bank controls, examine
UBS, also practiced entering false information into the bank's suspicious activity, and conduct incident-related activities
computers to hide the risky trades he was making, which eventually
cost the bank $2 billion.

At the heart of rogue trading (or other types of fraud) are human incentives: those who want to profit for personal gain or who enjoy
the thrill of excessive and unsanctioned risk taking, and those who are afraid to own up to losses. These incentives are reasons why
flagging rogue trading is a challenge in-house using traditional methods. Bank employees do not reveal problems early because they
are not incentivized to: they might get fired or lose their bonuses. Employers are not incentivized to be completely open with
regulators because of adverse effects on their business.

Algorithms and data-driven analysed by external teams of former traders, compliance staff, intelligence officials, and psychologists, to
a certain extend solve this incentive problem: systems alert to suspicious activity that is employee-agnostic, supported by an external
investigative team that is independent with minimal conflicts of interest.
A Re-evaluation of Artificial Intelligence’s potential?
Early this year, in a widely hailed new milestone for AI, Libratus, built by Carnegie Mellon University Professor of Computer Science
Tuomas Sandholm and his PhD student Noam Brown, won $1.5 million in chips after beating four of the world’s best poker players in
an extraordinary 20-day tournament.

Training a machine with incomplete, hidden and misleading information to win is significantly more challenging than constructing
layers of neural nets to beat humans at chess. Unlike chess where players see the entire board, poker players do not see each other’s
hands. From performing probability calculations to manipulating table image, poker is a game where the outcome is tied to players’
actions based on psychology and game theory. The ability to interpret an imperfect set of information and “bluff” is key to a winning
hand – and building this ability into artificial intelligence had proven to be elusive.

Libratus does this by self-learning: armed with massive computing power, it plays trillions of hands to refine its approach to arrive at a
winning strategy. Critically, Libratus does this overnight and repeatedly over the 20 days without needing to “take a break”; whereas
the poker pros face a very real physical challenge: they need to eat and sleep.
The success of Libratus is special. It challenges our preconceptions about the limitations of AI,
and takes us to previously unexplored possibilities: there is potential for applications from
negotiating trade deals to devising cyber security defense strategies to setting national
budgets – areas that we think of as strategic work with imperfect information.

But, AI successes such as this have also raised concerns. Aside from data protection issues in
Fraud Detection (will my personal investment data be anonymized for peer group profiling?),
or threats of surveillance in Market Behavioral analytics (will the storing of my phone and
electronic conversations be done in such a way that it meets legal requirements?), it is hard to
escape our nagging suspicions that AI will soon replace us.

The news that the world’s largest hedge fund, Bridgewater Associates which manages $160billion is extending AI beyond financial
trading to build “a piece of software to automate the day-to-day management of the firm, including hiring, firing and other strategic
decision-making” adds to the fears and insecurities felt by many of us.

Arguably, the examples provided here – Algo trading, Robo-Advisors, Fraud Detection, Market Behavioral analytics – do not eliminate
the human touch; AI merely collates data and draws out key information to allow for more efficient human decision making. An
Accenture survey of 1,770 managers across 14 countries concludes similarly: “AI will ultimately prove to be cheaper, more efficient”
and so will “free us from the drudgery of administrative tasks”, to allow us “to focus on things only humans can do.” However, some,
including the Futurist Ray Kurzweil, disagree and believe that what we think of as strategic work or even creative work can be
substantially overtaken by AI.

Perhaps, the real question is not if, but when: are we decades in planning for the arrival of full AI systems without human guidance? Is
it a quantum leap from today’s AI systems to performing strategic decision making? What research breakthroughs are required to
make these feasible? The evolutionarily path is unlikely to be a linear one, and the complexities of human activities mean that some
are easier to automate than others. But the rapid innovation of AI technologies mean that we should not dismiss the likelihood out of
hand. While the debate rages on, we can plan to adapt to AI’s transformational impact in our future lives. For the time being though,
we still hold some cards in our hands: there is no question that AI still needs our direction to set its objectives, programming,
algorithms, codes and ultimately, to turn it on.