 |
| |
 |
| |
|
| |
 |
Coauthored with Ted Goertzel
Webmind
Inc. was a company with a long-term mission: creating real
AI. But it also, like any other company, had the goal of
making money. And it wasn’t endowed with a large enough
amount of cash to enable a total focus on achieving real
AI, and making money after the long-term goal was complete
and a real AI was constructed. We had to build products
too, products we could sell in the immediate term, to increase
shareholder value and fund further research. From the very
start our idea with the company was to proceed on two tracks:
long-term AI research and short-term AI-based product development.
The most interesting product we made was our first one,
the Webmind Market Predictor -- a relatively simple software
system whose purpose was to predict the daily prices of
various currencies and futures, based on looking at the
prices of various financial instruments over the past few
days, and (this was the really innovative part) recognizing
patterns among the words, phrases and concepts used in the
recent news. MP (our nickname for the product) didn’t
use nearly all of the Webmind AI Engine, and the version
we finally productized didn’t rely on the main AI
Engine codebase at all, instead relying on simplified software
produced by combining more traditional statistical pattern
recognition based methods with some key features extracted
from the AI Engine (and specialized for the particular context
of financial prediction). In this case our would-be general
AI served the role of a prototyping framework in which ideas
leading to a specialized AI system were developed. Not a
realization of our grand goal by any means, but interesting
nonetheless.
It’s not conventional to think of finance as an important
aspect of the evolution of intelligence, of the path toward
the Singularity. But the fact is that as computers get smarter
and smarter, a lot of what they’ll be doing is managing
money. The finance industry and the military have long been
the two biggest users of AI. Nearly all of Hillis’s
Connection Machines were sold within these sectors. And
as much as the left-wing part of my mind doesn’t want
to admit it, the tech explosion that’s bringing the
Singularity about is essentially a capitalist phenomenon.
Money is not as interesting to me as computer technology
or neuroscience or genetics or ethics, but it’s certainly
right at the heart of what’s happening….
A question that naturally arises at this point is: If you
created such a great AI-based market prediction system,
how come you’re not rich? How come Webmind Inc. went
under instead of making it big on the markets? The answer
to that is simple, albeit a little embarrassing in retrospect.
The company made the mistake, in mid-1999, of back-burnering
the Market Predictor project, and focusing the attention
of its product development and marketing groups on a different
product: the Webmind Classification System, WCS, which divided
documents into categories automatically, and was sold to
a number of customers, mostly dot-com firms.
I was never terribly thrilled by this shift in product focus,
in fact I argued against it vehemently, but even so I was
never 100% convinced that the decision was wrong until events
had proved my intuition correct. It was our CEO Andy Siciliano
who made this decision, based on his years of experience
in the financial trading world, and his painful awareness
of the years it can sometimes take to turn an apparently
effective financial prediction system into a real moneymaker.
In 1999, the dot-com boom was at its peak, and the years
it could take to build up a fund based on a trading system
seemed like a slowpoke’s path to fortune, compared
to building some kind of Internet-focused firm and taking
it public. Lisa and I and the other founders always regretted
the way the MP was underemphasized within the company, but
I chose to ignore this worry and focus on the AI R&D
that was my main passion. Of course I also wound up spending
a huge amount of time dealing with the development and marketing
of WCS, even though I didn’t feel the product was
the direction we should be going in. And it turns out that,
although Webmind Inc., couldn’t build a business around
WCS, it’s possible to create a successful smaller
business around a similar piece of software. The guys from
the Webmind Inc. New Zealand office have started a new company,
ReelTwo, which is having some success with their Tui product,
a machine learning categorization system very much in the
WCS vein.
Anyway, Webmind Inc. dissolved in March 2001, due in substantial
part to the
failure of the WCS
business model. And ironically, in one of those twists so
typical of the business world, Andy Siciliano took the MP
code and began using it and further developing it on his
own, together with 5-6 former Webmind Inc. staff who had
been on that project during late 2000 (not including me).
The legal aftermath of the Webmind Inc. dissolution goes
on. But the bottom line technically is that the Webmind
MP does appear to work, and will potentially create a significant
profit for someone over the next few years, although neither
I nor the shareholders of Webmind Inc. are likely to be
in this particular category of “someones.” A
whole text could be written on the lessons about business
that I derived from this and other experiences with Webmind
Inc., but that would be a totally different book....
Investing
is a lot more popular these days than it was in 1997 when
I and my colleagues first designed the Webmind Market Predictor
-- let alone back when I was a kid, when it was something
hardly anyone ever talked about. Today everyone trades stocks,
and half the cab drivers of the world seem to believe they
know how to beat the market somehow -- maybe by surfing
the net to find a valuable piece of research everybody else
missed. Even if they can’t beat the market at least
they can gamble on it. When I lived in Vegas in the early
90’s I got really baffled at the reluctance of most
US states to legalize gambling – I figured it was
because they didn’t want to cut off the revenue they
made from lotteries. But now people have a new outlet for
the innate urge to gamble, and it’s perfectly legal.
The crash of the Internet bubble definitely put a damper
on the “man in the street’s urge to gamble on
stocks; but even post-crash, E-trade is doing a healthy
business.
Some people believe there’s really nothing but gambling
going on here. There’s a classic book by Burton Malkiel
called A Random Walk Down Wall Street. Malkiel argues that
a blindfolded monkey, picking stocks at random, could do
as well as the best professional financial analysts. Since
monkeys can't read, presumably the blindfold isn’t
much of a handicap ... the point is that, in his view, the
financial markets are not predictable.
Now one might think: “If everyone else in the market
is a blindfolded monkey, then, I should be able to make
a lot of money by entering the market and trading more intelligently.”
But as it is now, there are so many people trying to trade
intelligently, that any insight you might have has likely
been roughly simultaneously exploited by a dozen or a thousand
others, and hence "priced into the market." For
instance, if you have good reason to believe that IBM stock
is overvalued, and its price is going to drop soon, then
so do others, and these others will start selling IBM at
the same time as you. But this rush to sell IBM will make
the price drop. So you won't gain anything by selling it,
because it won't be overvalued anymore. This is called the
"efficient market hypothesis" -- the hypothesis
that there is no one on earth who can systematically do
better on the markets than a blindfolded monkey.
Until I began studying finance with a view toward Webmind
applications, this wasn't my view of the markets at all.
Rather, coming from a democratic socialist family background,
my perspective was that the markets were a way for the rich
and well-connected to profit at the expense of the working
and middle class. The markets, I was certain, were crooked.
Efficient, yes -- efficient at dumping money into the laps
of the richest and most dishonest individuals.
Well, I still see a certain amount of truth in this cynical
view. Insider trading is more common than is widely appreciated.
Maybe you have a friend who works at IBM and so you know
that their new products don't work. Then you know their
stock is going to drop and nobody else does. That's a way
to beat the blindfolded monkey for sure. (Poor innocent
beast!)
Another way, if you're an investment bank, is to create
a complex financial instruments in order to get around government
regulations regarding the kinds of investments mutual funds,
pension funds, etc., can taken on. Risky investments in
foreign currencies are cleverly disguised as investments
in US government bond derivatives, and so forth, so that
mutual and pension funds can "legally" buy them.
Deception is everywhere. And outright illegal activities
are also rampant. Mike Lissack, a friend of mine for several
years, achieved some notoriety as a finance-industry whistleblower.
Formerly an investment banker, he turned several large Wall
Street firms into the Federal Government for defrauding
the government out of billions of tax dollars. He had to
hide out in an FBI safe house for a while to avoid possible
assassination by his former employer.
And yet, in spite of all the crookedness, Wall Street also
contains a great number of very smart people, honestly trying
to beat the markets by one strategy or another. The motivation
is mainly greed, to be sure -- this is not an altruist's
game. But status is just as much a motivator, as well as
pure curiosity, and the challenge of the task. After all,
even if most trends and patterns are priced into the market,
what if there's one that nobody sees but you? Maybe you
can recognize from public data alone that the price of IBM
stock is going to drop, while no one else is clever enough
to see the truth. The efficient market view is basically
that no one is this much smarter or more knowledgeable than
everyone else, so that the rule holds 99.999% of the time
-- the markets can't be predicted. But no trader believes
this. They believe that it holds for almost everyone else,
but not for the best of the best. They're smart enough,
knowledgeable enough -- they can see the patterns that aren't
priced in.
In the efficient markets view, which is conventional wisdom
among finance professors in academia, it is not worth the
cost to invest in financial advice in the hope of "timing
the market" by buying stocks when they are low and
selling when they are high. Malkiel's book, which came out
in the early 70's and pushed the efficient market perspective
hard, actually had a big impact on the market. It helped
stimulate the development of index funds, which simply buy
a wide selection of the most reputable stocks - those that
are included in standard indexes such as Standard and Poor's
or Dow-Jones. Index funds have done very well for a long
time, although in the last year their performance has faltered
relative to stocks of smaller, hi-tech firms.
The people out there trying to beat the blindfolded monkey
fall into several different categories. First of all there
are "fundamental" analysts, who study the economic
fundamentals that determine an investment's true worth.
How much profit has it generated in the past, and how much
is it likely to generate in the future? How well is the
company managed? What are the future prospects for the industries
the company is competing in? These are, of course, difficult
things to analyze, but many analysts are well trained and
do a good job. This is what Lisa Pazer, my friend and Intelligenesis
co-founder, did for a living for many years, before she
quit Wall Street. She was a currency analyst. Lisa certainly
believed that she was recognizing genuine patterns in market
behavior, coming up with things that others didn't see.
But she traded herself for years, and never made it big.
Why were Lisa and her colleagues, in spite of their insights,
unable to beat the blindfolded monkey? In the efficient
markets view, the answer is that, the market has already
taken their analyses into account, before anyone can act
on the analyses to exploit them in a significant way. Thousands
of buyers and sellers, acting on this fundamental information,
generally arrive at a price that reflects each investment's
fundamental value. Or ... maybe they just weren't smart
enough!
Of course, rational analysis of companies' prospects is
not the whole story. Fundamental analysis clearly is not
the sole driver of the markets! Consider, for example, market
crashes such as those in 1929 and 1987. Suddenly, stock
prices go down by as much as a third. Certainly, the true
value of the companies' assets has not fallen that much
overnight. The explanation is that markets are also a psychological
phenomenon. They depend, not just on objective indicators,
but on how people appraise those indicators. So you get
periods of enthusiasm, when people build "castles in
the air," persuading themselves that the future is
unlimited for certain industries. And then the bubbles burst
– just like the moving Internet bubble I talked about
earlier.
Some financial analysts believe they can beat the market
by analyzing the financial trends. "Technical"
analysts try to do this by charting trends in stock prices.
They make graphs of trends in stock prices, and believe
that they can predict turning points by studying patterns
in the graphs. Some use more sophisticated mathematical
models rather than simple charts, especially now that computers
are available to do the computations. None of these technical
wizards has, however, established a really reliable track
record. According to the efficient-markets view, this is
because future trends in financial prices simply are not
correlated with short-term fluctuations in future prices.
One can predict long-term trends, within broad limits, but
this is basically what the fundamental analysts do, and
the results of their analyses are already incorporated in
today's stock prices.
In addition to the "fundamental" and "technical"
analysts, there are "behavioral" analysts who
do their best to follow trends in investor opinion. They,
in effect, try to psych out the market, anticipating when
the climate of opinion is about to change. This is a very
subtle field, depending on hunches and gut feelings, and
it is difficult to test statistically. Some of these analysts
have large followings, and make a lot of money selling newsletters
to people who believe in their theories. They have stories
to tell of great successes in predicting major turning points
in markets. But many of the most successful have gone on
to make dramatic bloopers. Just by luck, a certain number
of soothsayers are always going to be right, but relying
on the ones who were right in the past doesn't improve one's
chances in the future very much, if at all.
In the 1970's, a new group of financial analysts emerged,
called quantitative analysis. Unlike technical analysis
which usually involves recognition of fairly simple patterns
("after three consecutive peaks, expect a big fall"
and such), quantitative analysis uses highly sophisticated
mathematics to analyze the markets. Practitioners are called
"quants", or, more colorfully, "rocket scientists."
Rocket science is big business on Wall Street, and has led
to some huge successes and huge disasters. Last year, 1998,
Long Term Capital Management, a hedge fund run by some Nobel
Prize-winning rocket scientists, went under and lost billions
of dollars. They were trading in a way that was mathematically
guaranteed to succeed -- but it didn't. They lost anyway.
The real world did not agree with the assumptions of their
theorems. They held a number of investments that they believed
to be uncorrelated, but actually, when the Russian economy
crashed and a few other bad events occurred at the same
time, all of their holdings simultaneously tanked.
As even this brief overview indicates, financial analysis
is a very difficult and competitive field, and lots of clever
schemes have failed. The market mechanism itself seems to
guarantee that prices stick fairly close to their true value.
But, Lisa was convinced, based on her work as a fundamental
analyst, that there were patterns in the daily news that
a computer could exploit for market prediction. She had
been pretty good at picking up market-relevant news patterns,
but she felt a computer could do even better. It could read
more news than her, and study it more objectively. She convinced
me, back in 1997 when Webmind was just a rough design sketch
and a bunch of equations and concepts, that this was a good
initial Webmind application. This would be our first Webmind
"killer app" -- Webmind reading the news and predicting
the markets. It was wild, it was crazy, but it made sense.
Market prediction is a field where a small increase in intelligence
can reap tremendous rewards.
When
Webmind Inc. folded, the Webmind Market Predictor was still
in a testing phase. The test were going remarkably well,
however, and there was a lot of controversy within the company
as to whether the business should focus on MP, or on another
product that was also highly technically successful, the
Webmind Classification System (WCS; a tool for dividing
documents into categories). Ironically, the folks within
the company who were most familiar with the trading business
were eager to do with WCS, which played into the then-peaking
dot-com craze, in that most of the customers for WCS were
dot-com firms. Andy Siciliano, who had replaced Lisa as
CEO in mid-1999, was firmly in this camp. He was all too
familiar with the chancy nature of making money off the
markets. On the other hand, the folks in the firm with more
experience selling traditional software products were more
bullish on MP, because trading seemed to them a much simpler
and more direct way to make money – and they were
all too familiar with problems like a long sales cycle,
customer retention, and so forth. The WCS advocates won,
which turned out to be a very bad thing for the company
– which is easy to see in hindsight, but in foresight
none of us projected the precise timing of the dot-com crash,
nor its huge magnitude. We all knew the boom market couldn’t
last forever, but, a smaller crash would have been survivable
even with a business focused on WCS.
When the firm dissolved in March 2001, I hadn’t personally
been closely involved with MP for quite some time; I’d
been focusing on AI research, on endless business meetings
involving fundraising and WCS, and on the design of yet
another product, Webmind Search (a search engine). The MP
team, a handful of people spread across the US and Australia,
stayed together funded by Andy Siciliano personally. They’ve
been testing the system extensively since that point, and
Andy, with a fantastic finance-industry pedigree earned
through years as a trader and banker, is well-positioned
to transform an effective market-prediction system into
a money-making trading firm. So I suspect we haven’t
heard the last of Market Predictor. (And, incidentally,
MP and Novamente are not the only surviving offshoots of
Webmind Inc.; the core of the WCS team, the Webmind Inc.
New Zealand office, has formed a new company called ReelTwo,
selling a new text categorization product.)
In scientific terms, what the success of Webmind MP –
and the handful of other advanced AI trading systems that
are out there – shows is that the financial markets
are not truly efficient. What they are is almost efficient
with respect to human intelligence. There are enough smart
humans trading the markets, that almost any inefficiency
detectable by a smart human will be immediately detected,
and priced in. But, there are inefficiencies in the market
that are not easily detected by the human mind, but are
nonetheless real. Webmind MP, as simple as it is to the
full Webmind AI Engine or Novamente, is a non-human mind
of a sort, and it can detect different patterns in the market,
thus exploiting inefficiencies that humans cannot. Specifically,
by reading the news and using concepts it extracts from
the news to predict the markets, Webmind MP is detecting
trends in human mass psychology that humans are not detecting.
This is a fascinating accomplishment in itself, even if
you have no interest in using it to make money.
There are dozens of AI products aimed at financial prediction
-- using neural nets, genetic algorithms, and expert-system-type
rules -- but most of these offer only small performance
gains over the blindfolded monkey. This because they are
really not all that intelligent. As artificial intelligence
technology develops, we will see more and more situations
like the one we currently have with Webmind MP -- exploitation
of patterns in the market that no one has detected before,
because there never before existed a mind with the proper
orientation. Of course, if everyone started using tools
Webmind MP to predict the markets, then Webmind MP's intuitions
would become priced in, and the inefficiency would be gone.
You'd need a new version of Webmind MP, or a different kind
of AI, to gain an advantage. In the financial markets of
the future, the spoils may to he or she who can develop
a better, more distinctive, artificial brain.
The thing that’s unique about the Webmind MP, as opposed
to other financial AI products, is its ability to analyze
and synthesize both quantitative and qualitative data. It
reads both text and numbers. In addition to following statistical
trends and synthesizing the results of nonlinear predictive
algorithms and standard financial indicators, Webmind reads
the news, just as human analysts do. We simply feed in text
from readily available financial news services. The system
reads this news, not in an undirected, musing kind of way,
but with a particular financial data set in mind, say the
Dow Jones. It constructs concepts that capture themes in
the news which are correlated with what the market is going
to do the next day (or the next hour, or 2 weeks later,
or whatever). The financial meaning of the text is thus
boiled down to a collection of numbers - one for each concept
extracted, representing the relevance of that concept to
the text on a certain day. The numbers corresponding to
the concepts can be computed anew every day, or even more
often, and used for financial analysis purposes just like
numbers coming from any other source. These numbers, representing
the relevancies of text-derived concepts to the news at
a particular time, are what we call text indicators.
The extraction of text indicators is the crux of Webmind
MP's financial intelligence. It relies on Webmind MP's ability
to intelligently judge relevance, which draws on all of
Webmind MP's abilities at reasoning, language understanding,
conceptualization, and so forth. But text indicator extraction
is not the end of the story. In addition to the extraction
of financially relevant concepts from news, there is an
additional process of learning optimal trading models for
particular financial markets. Just knowing the news concepts
that tend to correlate with a certain financial market (say,
knowing that trouble in foreign countries tends to drive
the Dow) doesn't tell you enough to make accurate predictions.
You have to get at the nonlinear interrelations between
news concepts and numerical patterns in the data. This has
to be done differently for the Dow, for IBM stock, for the
Yen, for 30 year bonds, and so forth. For each market, Webmind
derives a trading model that embodies the best way of incorporating
text based information into decisions about that particular
market.
Webmind's trading models are what computer scientists call
"Boolean automata," simple logical decision rules,
just as Webmind uses for making any kind of decision. They're
basically the same kind of rules that are used, within the
natural language system, to decide which sense of "Java,"
is intended in a sentence (the computer program, the island
in Indonesia, or a copy of coffee). More generally, they
are the rules that Webmind follows for learning abstract
concepts and categories.
For reasons of efficiency, in building Webmind MP we adopted
a special and simplified format for trading decision rules;
a format derived from the prior work Jeff Pressing, an Intelligenesis
co-founder. Jeff is yet another super-brilliant multitalented
scientist: physics PhD; acclaimed jazz pianist, West African
drummer, multi-instrumentalist and composer; currently working
as a psychology professor … and does financial prediction
on the side. For the few years prior to the founding of
Webmind Inc., Jeff was trading the Australian bond market
for a group of Australian investors, using trading rules
of his own invention and making a fair amount of money.
Jeff's rule format doesn't make Webmind MP perform any better
than it would if it used its default decision rule module;
but it does make it learn faster. In the current configuration,
the system would take about an hour to learn a trading model
using the generic decision module, as compared to about
one minute using Jeff's streamlined framework. This is the
kind of tradeoff that you face all the time doing AI engineering:
the more specialized you get, the better your performance
in one particular domain, but the less generalizable the
performance is to other domains.
Initially, when Jeff and I developed this approach, Lisa
didn't see why this decision-rule-inferring phase of the
process should be necessary. "Why," she asked,
"isn't it enough just to determine the concepts, occurring
in news, that drive the markets? People are sheep,"
she said, “they just follow the herd.
My answer was simple: "People may be sheep, but they're
not retarded sheep."
The serious answer, of course, is self-organization, nonlinearity.
The markets are a mind of a sort -- which means that even
when you know what motivates them, there is a great deal
of subtlety to exactly how this motivation occurs. Just
because the Dow tends to be driven by trouble in foreign
countries, this doesn't mean that every time there's trouble
in foreign countries, the Dow's going to jump. The conditions
have to be right for this connection to manifest itself.
An analogy is, suppose you've figured out that a given person
is a sucker for beautiful women. This doesn't mean that
every time the guy sees a beautiful woman, he's going to
react in a certain way. If you want to be sure of his reaction,
the conditions have to be right. You want to catch the guy
at a good time of day; you want to catch him when it's been
at least a few hours since he was with the last beautiful
woman, etc. People, and markets, are predictable, but not
linearly so. They react to complex combinations of stimuli,
spread out over space and time. It takes an intelligent
system, like a human or a Webmind AI Engine or Novamente,
or at least a Webmind MP, to predict what they're going
to do with reasonable effectiveness.
Table 2 below shows a small selection of the hundreds of
amazing results we obtained in the early days of testing
Webmind on financial analysis problems. In these simple
experiments, we asked Webmind to learn optimal trading models
for five major markets, first with and then without news-derived
information. When it included the information derived from
news archives, the performance increased tremendously, often
by a factor of 3 or more.
|
Market |
Entry
type |
profit/year
without text |
profit/year
with
text |
Increase
in profit due to use of text |
| DJIA |
Long |
63.8% |
172.3% |
170.0% |
| Eurodollar
futures |
Long |
22.9% |
134.4% |
486.9% |
| yen
futures |
Short |
35.4% |
118.3% |
234.1% |
| US
T‑bond futures |
Short |
7.8% |
95.8% |
1128.2% |
| Cocoa
futures |
Short |
19.9% |
30.6% |
53.8% |
|
| Table 2 - An
illustration of the power of text based information
in market prediction on 5 markets |
These
results are old ones, obtained from “simulated backtesting”
of the system; there are much more current results involving
actual trading. But those are being held as proprietary
by the current Webmind MP team; the reason I can show these
results here is that they were presented by Jeff and myself
in a published conference paper in 1999.
These results are dramatic in the context of trading, and
they’re also of fairly obvious importance outside
the financial arena, because the dynamics underlying Webmind
MP 's performance on these sample tasks are the same ones
underlying more general aspects of the Webmind AI Engine
and Novamente. In addition to demonstrating that Webmind
MP can produce highly profitable trading systems using textual
information, these tests also show something more general:
that Webmind MP can meaningfully extract concepts from large
amounts of text data, relate these concepts to trends in
numerical data, and use the relationships it has inferred
to provide highly significant actionable information. This
capability is of tremendous value beyond the world of finance,
in areas such as risk assessment, demand planning, supply
chain management, enterprise management, document retrieval
and analysis, primary market analysis, medical diagnosis
and research -- the list could be continued almost endlessly.
Wide implementation of this kind of financial prediction
technology – which is all but inevitable, given a
little time -- will have a globally dramatic effect. The
financial component of the World Wide Brain will become
vastly more intelligent than it is right now. I would hate
to see the World Wide Brain develop with an overly financial
bias -- the finance industry, for all its breadth, does
after all represent a fairly narrowly biased view of the
human race and all the information it has to offer. But
the financial markets are already a globally integrated,
perceiving and acting self-organizing system, and so they
are a natural place for Internet intelligence to start.
The
precise mechanisms underlying Webmind's text-based market
prediction are, obviously, proprietary. They're also patent-pending:
Lisa and I applied for a patent for the details of this
process in mid-1998. But the basic character of the process
is not a secret, as it's nothing but Webmind intelligence,
applied to one particular domain. What I’ll describe
here is a very early implementation of Webmind MP, back
when market prediction was done in an old version of the
Webmind AI Engine, before it was split off into a separate
product on its own. This Webmind MP didn’t have a
version number, so I’ll just refer to it as the “early
Webmind MP.”
Recall the basics of Webmind architecture. Webmind, internally,
consists of a collection of software objects called "nodes,"
each of which contains links to other nodes, representing
inter-node relationships. Some nodes contain raw data such
as text or numerical time series; others are more abstract
and consist entirely of links to other nodes. A Webmind
node is more like a "neuronal module" in the brain
than it is like a single neuron. In Webmind , unlike in
a neural network, link construction is carried out by a
variety of intelligent software actors, and nodes and links
are frequently created and destroyed as part of the learning
process. Webmind 's internal intelligent actors use a variety
of techniques such as genetic algorithms and statistical
language processing.
In the early Webmind Market Predictor system, different
types of nodes were used for representing different types
of data. The node types most directly relevant to simple
financial applications are:
·
DataNode, which refers to a numerical data sets, e.g.
financial time series
· TextNode, which refers to a text document, e.g.
a market news report for a particular day
· TimeSet, which refers to a series of time-indexed
nodes, e.g. a series of market news reports over several
years
· ConceptNode, which refers to a concept either
extracted from text or learned by Webmind
· TradingRuleNode, representing a financial trading
rule
· TradingSystemNode, representing a trading system,
which is a collection of trading rules
The
process of learning “textual indicators” to
aid in market prediction, trading and analysis was, in this
early Webmind MP version, a natural outgrowth of Webmind
's intelligent self-organizing dynamics. Consider, for simplicity,
the case where the market being trading is the Dow Jones
Industrial Average, and the text being used is market news
which is issued on a daily basis. In this case the central
data structures are a DataNode pointing to the DJIA, and
a TimeSet pointing to the series of market news articles.
The text indicator learning process is carried out by an
actor associated with these two nodes. The goal of this
actor is to isolate a ConceptNode, representing one of Webmind
's internal ideas, with the property that the relevance
of the ConceptNode to the TimeSet, on a given day, is a
useful indicator for trading the data in the DJIA DataNode
on that day. Having carried out this learning process, it
then uses the ConceptNode it has isolated to create a new
"text indicator" DataNode, each entry of which
indicates the relevance of the ConceptNode to the market
news TimeSet on a certain day. This text indicator DataNode
can then be supplied to the user as, quite simply, a series
of numbers. Each day, on reading the news, Webmind MP can
supply a new value for the text indicator -- just as, on
seeing the DJIA itself, it can supply new values for numerical
indicators such as moving averages and the like.
The text indicator obtained by this process of internal
conceptualization can be used for a variety of different
purposes. It can be used by humans to form their own intuitive
judgements, or incorporated in any computational trading
framework that is sufficiently flexible to incorporate arbitrary
numerical indicators. Most of our practical financial prediction
work, however, involved incorporating text indicators into
Webmind MP's own intelligent trading processes, which are
ideally suited for incorporating text indicator information.
A TradingRuleNode, in the early Webmind MP, encapsulated
a trading rule, which was a logical combination of information
deriving from a numerical predictor and a number of indicators,
some numerical and some textual.
An indicator, generally speaking, provides a piece of information
about the state of play at the time computed. There are
hundreds of different indicators in common use in financial
analysis today. For example, an indicator may be the value
of the market series at that time, or the 5-session variability
at that time, the "Relative Strength Index" of
another data series at that time, or a text indicator --
the relevance of a particular concept within Webmind MP
to the news at a particular time.
A predictor, on the other hand, doesn’t merely give
you some information about the market -- it tries to tell
you what the market is going to do. It takes in the past
history of a market series, and perhaps other past information,
and predicts the next value of the market series. The early
Webmind MP used various types of predictors, based on such
approaches as linear and nonlinear regression or pattern
matching (e.g., based on fit to past history of the same
market series, or changes in the series, or weighted historical
combinations of various indicators). The parameters characterizing
each predictor can be optimized by Webmind MP to produce
maximally effective prediction. These parameters can then
be stored for later recall. More sophisticated prediction
in a certain market can be addressed by an optimally evolved
combination of one or more predictors in that market, indicator
or predictor information from other markets, and Webmind-MP-extracted
text-based indicator series. Prediction can be tested by
discrepancy between the actual market outcome and the prediction,
or by a what-if analysis of the impacts of decision-making
recommendations.
So, basically, once it had read text and made text indicators
from them, the early Webmind MP used evolutionary methods
to automatically "evolve" trading rules and trading
systems utilizing the predictors and indicators it has at
its disposal, optimized for performance on particular financial
instruments. The evolution of trading rules provides a quantitative
way to test the ability of Webmind to find exploitable windows
of enhanced predictability for financial decision-making.
Evolved trading rules may be either long or short. Trading
rules may then be combined to produce trading systems, which
consist of at least one long rule and at least one short
rule. There are various technical criteria telling a rule
when to exit the market after it’s identified a good
time to enter.
A simple example of a trading rule this Webmind MP might
have used is:
IF
Simplex predictor predicts > 0.25 % rise AND
Fast Stochastic Indicator (31 sessions) > 36 AND
Relevant Textual Indicator < .15
THEN ENTER LONG
EXIT after 1 session
This
rule uses a predictor, one numerical indicator and one textual
indicator; it combines these using numerical inequalities
(involving numbers Webmind MP thought up) and logical operators
(AND in this case). This rule finds a good time to enter,
and then exits after one session, banking its winnings.
Basically all this work just confirms the wisdom that, while
financial markets are complex and difficult to predict in
general, there are windows of enhanced predictability, and
if you’re smart enough, you can find them. With its
text indicators and its evolutionary learning of rules that
incorporate them, Webmind seems to be smart enough.
So
far I’ve been fairly mysterious about one very crucial
point in the "Webmind MP as financial guru" story.
What exactly are these "concepts" that Webmind
MP derives, by reading the news with financial data in mind?
This is the most fascinating part of the whole tale! These
are not simple, sensible English concepts. It's not like
the Yen futures jump every time the word "hurricane"
is mentioned, or every time a story about Japanese industry
appears on the front page of the New York Times. The concepts
that drive the market are a bit subtler and more abstract
than this. Within the early Webmind MP, they were represented
as abstract data structures -- ConceptNodes. They are associated
with various words and documents and numerical data sets,
but they are not truly described in any way except as abstract
Webmind-internal data structures. Just as, while we can
try to describe our intuitions about the world in words,
we can never quite capture them. A good boxer, or financial
trader, or mathematician can try to explain the intuitive
ideas that guide his or her work, but the articulation never
quite matches the reality.
For the period 1996-1997, for instance, the most prominent
concepts driving the Dow and the Yen had to do with trouble
in Asian economies, and Asian banks particularly. "Asian
bank trouble" as a focus of the news was highly predictive
of downturns in the Dow. One useful concept had to do with
the relation between the British government and the Asian
financial situation. Another had to do with European unification.
But the important thing to remember is that the concept
that Webmind MP found to be related to the market wasn't
just "European unification," itself. It was, rather,
a certain slant on European unification, a certain sense
it got from the news, that had to do with European unification.
These are fairly concrete concepts. But the longer the time
scale Webmind MP studies the news over, the more abstract
its concepts become. Optimal prediction uses a combination
of concrete, short-term-relevant concepts, with longer-term
concepts getting at more fundamental underlying patterns.
One of the most interesting concepts to come out of long-term
analysis is general reluctance to take a stand. This is
an internal concept which is activated whenever Webmind
MP is reading news articles in which the author is apparently
unwilling to state his own opinion, and is instead citing
other sources, discussing the opinions of other analysts,
etc. This concept is a good predictor, not specifically
of increase or decrease in the market, but of increasing
volatility in the market. When people start passing the
buck to others, this means things are about to go nuts.
Intuitively, it seems that the long-term patterns Webmind
<P picks up in its news-based financial analysis are
archetypal, whereas the short-term patterns are situational.
For example, we haven't actually seen this one yet, but
I am quite confident that as we do more analysis of long-term
news reports and market data, we're going to find the Good
Guys/Bad Guys archetypes popping up. These may not come
up explicitly, but they will come up as tones of expression,
as general moods of the collective mind, just like the "reluctance
to take a stand" collective mood that we've detected
in our current experiments.
Whatever else it may be, Webmind MP is an outstanding tool
for studying collective mind. It is a pattern-ometer, and
an archetype-ometer. By looking at what humans write, with
a specific goal in mind, it gets at the subtle patterns
underlying what humans are saying and thinking at a particular
point in time. This is the kind of thing that we have always
sensed, but never before measured. Because this is the kind
of thing it takes an intelligence to measure ... intelligence,
but not necessarily human intelligence!
|
|
|
