快猫短视频

IT鈥橲 THURSDAY 23 October and it鈥檚 hell. The Hong Kong stock exchange has just
taken a nose dive and brokers everywhere watch in a mixture of fear and
fascination as the rest of the world鈥檚 markets follow suit. Should they sell now
and cut their losses? Should they buy while prices are crashing down in the hope
that the markets will pick up? Or should they hold their breath and wait for the
storm to play itself out? Only good instincts will save them from doom.

To varying degrees, 鈥渢he hunch鈥� is what life as a dealer or analyst is all
about. Even in less extreme conditions, they need to ask themselves a thousand
questions a day. Is the latest interest rate hike good or bad? Is the market
overvalued? How will a company鈥檚 share price react to a financial scandal?
Hardly surprising then that this volatile, nervy world would be deeply resistant
to any attempt to turn those well-honed hunches into smart software, capable of
making decisions to buy or sell. And yet, slowly, reluctantly, cautiously, the
financial world is starting to put its faith鈥攁nd loads of money鈥攊nto
artificial intelligence.

To be fair, we鈥檝e been here before. Finance houses got excited about AI in
the late 1980s. But on that occasion, early success was swamped by failure.
Today鈥檚 excitement surrounds groups that have persevered with AI ever since, not
sticking with one technique but combining different methods. Some of the world鈥檚
wealthiest and most conservative institutions are now handing their money to the
machines these groups created. In October, the Boston-based State Street Global
Advisors, the third largest investment management company in the US, took over
Advanced Investment technology (AIT) of Florida, a pioneering designer of
electronic analysts. So mainstream finance houses are moving in on AI.

In London, Pareto Partners manages about $20.5 billion for government
and corporate pension funds. When it was set up in 1991, its director of
research, Ron Liesching looked for a partner with whom to develop technology to
help make complex investment decisions. He already knew how hard it would be to
build this technology. 鈥淭here鈥檚 a lot of noise in the data,鈥� he says. 鈥淎nd
you鈥檝e got to get the job done reliably. If you鈥檙e wrong, you鈥檙e gone.鈥�
Realising that financial markets are like war zones, Liesching cast his eye
towards the military鈥檚 research labs. At Hughes Electronics, a California-based
contractor to the US Department of Defense, Liesching found what he was after:
an expert system for extracting and automating the thought processes of
battlefield officers.

An expert system is a program that encapsulates the specialist knowledge of
an expert, which others can use to take the expert鈥檚 place. The hardest part of
building such a program is converting that knowledge into software. People are
often not aware of the reasoning they use to come to a decision, or find it hard
to quantify their judgments. How, for example, does a tank commander make the
trade-off between achieving an important military objective and the heavy troop
losses this action could lead to? Even if the reasoning can be worked out,
there鈥檚 the problem of translating terms such as 鈥渢rade-off鈥�, 鈥渋mportant鈥� and
鈥渉eavy鈥� into computer code. To extract and define these sorts of terms and
thought processes, developers of expert systems have evolved a set of techniques
they call knowledge engineering.

Most existing expert systems tend to operate in areas where problems are
solvable by straightforward reasoning, such as backward or forward chaining.
Backward chaining is where the expert starts with a hypothesis and tries to work
out whether it is true. For example, a doctor who thinks a child has measles
might check the patient鈥檚 symptoms to decide whether this hypothesis is true. In
forward chaining, the expert starts with the symptoms and works forwards to
conclude that the child has measles. Unfortunately, financial trading is often
much more complicated, involving many variables and different ways of reasoning,
says Liesching. Conventional expert systems proved too inflexible for analysing
global bond markets.

Evolving intelligence

At Hughes, researcher Charles Dolan used a mixture of old and new AI ideas to
develop the system that caught Liesching鈥檚 eye. Expert systems take a 鈥渢op-down鈥�
approach, in which scientists attempt to analyse an aspect of intelligence and
then write a program to simulate it. But in recent years, scientists have
experimented with 鈥渂ottom-up鈥� approaches: they build a model of an organic
structure or process, which they then set running to 鈥渓earn鈥� or 鈥渆volve鈥�
intelligence.

Neural networks are the best-known example of this way of doing things.
快猫短视频s 鈥渢rain鈥� a neural network by presenting it with examples of a
problem鈥攕uch as identifying a car in a digital image of a street鈥攁nd
adjusting the connections between the neurons until the network produces the
correct output鈥攕ay, drawing an outline round the car鈥攆or every
example
(see Diagram).
This 鈥渃onnectionist鈥� approach allows
scientists to simulate an intelligent process even when they don鈥檛 understand it
in a way that can be captured in equations.

Dolan鈥檚 innovation has been to use similar thinking to generate sophisticated
expert systems. Rather than attempting to reduce a person鈥檚 expertise to a
single orderly list of facts and rules, Dolan creates a complex network of
different thought processes that interact to reach a decision, in much the same
way as neurons in a network do. He has developed a suite of programs, called
Modular Knowledge Acquisition Toolkit (M-KAT), that mimics different ways of
thinking to help extract and model an expert鈥檚 knowledge.

With his toolkit, Dolan modelled the thinking of Christine Downton, Pareto鈥檚
bond market expert. Downton is a star analyst with 20 years鈥� experience as an
academic, fund manager and banker. Over a period of 18 months starting in 1993,
she flew regularly to Los Angeles for brain-dumping sessions with Dolan.

The most difficult part of the task was capturing how Downton homed in on key
information hidden in a mass of economic data鈥攁n ability called 鈥渇eature
extraction鈥�. At a glance, Downton can tell if an economic measure is 鈥渉igh鈥� or
鈥渓ow鈥�, or if a price has moved 鈥渁 lot鈥� or 鈥渁 little鈥� and so on. The system had
to 鈥渓ook at a particular number and come up with exactly the same assessment as
I would鈥�, says Downton. To make it do this, Dolan brought fuzzy logic into
play.

Fuzzy logic allows computers to cope with these imprecise sorts of
assessment. The trick is to ignore the actual value of, say, a financial
variable. Instead, fuzzy logic programs look at where the value stands within
its range of potential values. So, for example, if the current inflation rate
falls within the lowest 20 per cent of its potential range it would be
considered 鈥渓ow鈥�, while if it is within the top 20 per cent it would be seen as
鈥渉igh鈥�. The rules of inference of fuzzy logic work with these 鈥渇uzzy鈥� values to
reach conclusions where there are no precise answers.

Dolan boiled down Downton鈥檚 expertise to some 2000 rules. The system, called
the Global Bond Allocation Strategy, runs on an Apple Macintosh. Every month it
pulls in around 800 items of economic information on 14 bond markets, taking
them directly off electronic data feeds. This information includes countries鈥�
public sector borrowing requirements (PSBR), unemployment figures, inflation
rates, money supply figures and so on.

Undervalued

Dolan鈥檚 system automates the process that Downton would use to decide which
bonds are overvalued and which are undervalued. The computer can look at a bond
price and say to itself the equivalent of: 鈥淢mmm, this looks high given the
country鈥檚 interest rate, I wonder what direction other economic indicators, such
as PSBR, are pointing?鈥� Towards the end of the process, the system compares its
new analysis of every bond with stores of typical cases. If the situation of any
bond or group of bonds resembles a past situation, it will consider the way the
market developed on that occasion. Finally it recommends a series of bond
trades.

Financial regulations mean that Downton must keep an eye on the system, but
she never interferes with its decisions. Nevertheless, for the final step, the
machine has to hand over to a person. Physically trading the bonds is still the
exclusive preserve of humans, so its recommendations are passed to a trader to
strike the deals.

Humans also do the physical trading at AIT in Florida, which deals on
American stock markets on behalf of pension funds. But, once again, computers
decide which trades to make. Every week AIT鈥檚 machines sift through 24 million
pieces of information and evaluate 3000 stocks, searching for clues as to which
will rise in price and which will fall. As well as stock prices, they take
account of details such the earnings estimates of quoted companies, their cash
flows, and current book values鈥攖heir assets minus liabilities.

The enormous volume of data was not the only barrier that needed to be
overcome in designing the system, says Dean Barr, president of AIT. Factors such
as earnings, cash flow and book value interact with one another and affect the
stock price in 鈥渘onlinear鈥� ways. In other words, a combination of tiny changes
in some of the indicators can trigger a massive shift in the stock price.
Conventional mathematics cannot deal with systems that exhibit this kind of
behaviour鈥攚hich is where the new AI comes in.

Just like financial variables, the neurons in a neural network interact in
nonlinear ways. So when AIT looked for ways to simulate the relationships
between the principal financial variables and stock price, neural networks were
a natural choice. Once the network has been trained to associate past financial
data with ensuing stock prices, it should be able to calculate the likely stock
price from any new values for the variables.

But creating a neural network system is not just a matter of shovelling in
any old data and waiting for a forecast to pop out the other end. Selecting the
right network architecture and variables to feed in, and choosing which
historical data to use to train and test it, all influence the success of the
analysis, says Barr. The chosen combination is a closely guarded company
secret.

The kind of pitfalls that exist is exemplified by the military researchers
who tried to teach a neural network to identify tanks in a landscape. They
trained the network on photographs of scenes in which tanks were present or
absent. When they tested the network on a new set of images it performed
brilliantly. Only then did the researchers realise that all the images with
tanks had been taken on a bright day, while scenes without tanks had been
photographed on a dull day. The network, it turned out, had learnt to
differentiate between dull and bright skies. It didn鈥檛 look for tanks at
all.

To avoid this kind of blunder, AIT carves up its data on stocks in several
different ways and runs the learning and testing routines several times. To give
added assurance, AIT also makes alternative forecasts with genetic
algorithms鈥攑rograms that can 鈥渆volve鈥� a solution to a problem, starting
from a set of random solutions. A genetic algorithm treats a problem as an
鈥渆nvironment鈥� and solutions as 鈥渓ife forms鈥� struggling to flourish. The program
forces the solutions to evolve over many generations until it identifies the
鈥渇ittest鈥� life form for the environment鈥攚hich is the optimal solution.

Fitness test

Suppose the problem is to find which set of economic variables best predicts
future movement of a company鈥檚 stock. The algorithm first generates a set of
solutions, each of which includes some or all of the variables, randomly
weighted and added or multiplied together in different ways. Next, it subjects
each solution, encoded as a string of 1s and 0s, to a 鈥渇itness test鈥� to measure
how well it tracks the stock price movements.

Unsuccessful solutions are killed off, while more successful ones survive
into the next generation. The algorithm enhances the evolutionary process by
allowing some surviving solutions to 鈥渕ate鈥�, by swapping sections of code with
other successful solutions. It also 鈥渕utates鈥� them by changing a few 1s into 0s
and vice versa. The new set of solutions then undergoes the fitness test again,
and the process is repeated over and over until an optimal solution evolves.

For every stock that AIT trades, it has built an electronic
analyst鈥攅ach made up of several neural networks and genetic algorithms.
These elements work independently, forecasting future stock price movements by
analysing past movements or by checking sets of information about the company
itself. Near the end of the process, the electronic analyst holds a 鈥渃ommittee
meeting鈥� at which it combines the various elements鈥� forecasts to produce a
recommendation to buy, hold or sell.

The systems at Pareto and AIT are the state of the art in applied artificial
intelligence. The question that really matters, though, is how well they
perform. Pareto鈥檚 Global Bond Allocation Strategy began operating in September
1994 and has been coming in at around 4.1 per cent above the bond market
benchmark鈥攁 reference point based on a portfolio of key bonds. These
rewards are modest, but then Pareto has tuned its system to a low-risk strategy
because of the conservative nature of its clients. The returns have been good
enough for Pareto to attract an increasing flow of funds from clients such as
the respected Mellon Bank, based in Pittsburgh, Pennsylvania. In Florida, AIT鈥檚
longest-running portfolio of stocks has shown an average annual return over
three years of 4.5 per cent above the Standard & Poor 400 stock index
benchmark鈥攁 selection of 400 leading US stocks.

But are the machines doing any better than the humans they are replacing? In
terms of performance, the Global Bond Allocation Strategy is in the top quarter
of investment management funds in the same market. Downton thinks her electronic
clone has some significant advantages, particularly because computers don鈥檛
suffer from stress or emotional disturbances. 鈥淓motions distort people鈥檚
rational judgments,鈥� she says. 鈥淭here鈥檚 a fear factor in managing money. People
tend to make mistakes when they鈥檙e losing money. They also make mistakes when
they鈥檝e made money, because they get big-headed.鈥�

Liesching sees similar advantages. Expert systems do not suffer from
鈥渃ognitive biases鈥�, he says. Humans can become fixated on one variable, or get
hung up on the most recent piece of information they receive. 鈥淚nvestors
frequently focus on one or two factors at a time, rather than assessing the
whole environment,鈥� he says. This can mean they miss important developments
elsewhere.

There is also a difference in brute processing power. 鈥淓very weekend our
electronic analysts are evaluating 3000 stocks against new information using
adaptive techniques,鈥� says Barr. 鈥淭hey do it with so much speed and efficiency I
don鈥檛 think people can compete.鈥�

Act of faith

But if computers are so fast and efficient, why isn鈥檛 everybody using them?
The two big reasons are time and money. Developing AI for the financial sector
has taken determination and long-term commitment, says Barr, and in the
financial sector the latter is in particularly short supply. 鈥淭he typical Wall
Street attitude is if you don鈥檛 get quick results you drop it,鈥� Barr says. AIT
set out to do things differently. 鈥淥ur company was created with the concept that
we would see this thing through to the finish line,鈥� Barr proclaims. And it鈥檚 an
expensive business. Liesching reckons it costs at least 拢1 million to test
a new expert system or neural network.

But many analysts and traders are still suspicious about AI, believing that
its success will be short-lived. They say the reasoning underlying today鈥檚 AI
systems is based on data from a particular set of market conditions. As the
market moves, those conditions will inevitably change, leaving the machines high
and dry.

Barr rejects such pessimism. Human analysts keep a close eye on their
electronic counterparts and constantly tweak them, so they learn all the time.
鈥淲e are not defying the laws of physics,鈥� says Barr, but evolving techniques
that analysts have applied for decades. Downton regularly analyses the bond
markets to see if they have changed in some fundamental way that would undermine
Pareto鈥檚 expert system. So far they haven鈥檛.

Today, an increasing number of big financial institutions, including the
Banque Nationale de Paris and the World Bank, are experimenting with AI. State
Street Global Advisors has set up its own AI laboratory with Barr at the helm.
Other firms using AI include the New York-based bank D. E. Shaw and the
Prediction Company of Santa Fe, New Mexico, which is working for the investment
bank SBC Warburg Dillon Read. But these are reluctant to reveal what they鈥檙e up
to.

If interest in AI continues to grow, perhaps analysts have good reason to
worry. AIT鈥檚 machines do the jobs of hundreds of humans. According to Liesching:
鈥淧eople in finance are generally overpaid, under-qualified and there are too
many of them. Whoever can replace these people with machines will win鈥攅ven
if the machines are only half as good鈥攂ecause they can work 24 hours a
day.鈥� The systems at Pareto and AIT are in the vanguard of a process of
automation that will rock the financial industry to its foundations, says
Liesching. Anyone tempted to reject this as just another wild claim of AI
enthusiasts should take note: the message is coming not from computer scientists
in ivory towers but from people who are prepared to put big money where their
machines are.