Artificial Intelligence

and Value Investing

Korok Ray

M
Korok R ay achines today can drive your their ability to reason, not just use statistical
is an associate professor car, pilot your airplane, tran- techniques.1 The push for a viable technology
of accounting in the
scribe your emails, vacuum that can mimic and even replace the human
Mays Business School
and Director of the Mays your f loors, search for your value investor will have long-term conse-
Innovation Research spouse, educate your children, power your quences for both AI itself and the efficiency
Center at Texas A&M home, and organize your factory. The trends of capital markets. I will begin by discussing
University in College in technology and artificial intelligence (AI) the modus operandi of AI. I will then discuss
Station, TX. over the last 50 years will continue, and and recast value investing in terms of this
korok@tamu.edu
machines will only expand in power, intelli- method. Finally, I will mine the AI scien-
gence, and scope. The question remains: Can tific literature for the most relevant bodies
they allocate capital? In particular, can they of knowledge that can help implement value
do so in the classic style of a value investor investing.
like Benjamin Graham or Warren Buffett?
Yes. This kind of AI differs from the A SKETCH OF AI
algorithmic trading currently in vogue in
financial markets. Most algorithms today Most are familiar with programming
are based on large-sample analysis that uses languages. One layer of abstraction above the
machine learning techniques from AI to pro- programming language is the algorithm, a set
cess information on thousands of securities, of instructions given to a machine written in
using millions of data points as input. These pseudocode, a shorthand that expresses the main
algorithms usually do not take concentrated ideas in mathematical logic. Algorithms can
positions, nor do they hold these positions be expressed in multiple languages (e.g., C++,
for very long. For machines to buy and hold Python, and Java) because the language
over long horizons, they will require a more serves to implement the fundamental ideas
radical and advanced form of AI that is now from the algorithm. There is an even higher
emerging. level of abstraction above the algorithm: the
This article provides speculation on representation of the problem. Every algo-
what this kind of intelligence looks like rithm must work within a representation, but
and how exactly it may come to be. It seeks there may be several ways to represent each
to connect two broad and disparate areas: problem (Lin [1965]).
the sprawling domain of AI and the more For example, consider the classic
old-fashioned style of value investing. The 8-puzzle in Exhibit 1. The user manipulates
crux of the analysis rests on computers and a 3 × 3 grid by moving tiles into a single

Spring 2018 The Journal of Investing    21

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Exhibit 1 REPRESENTING THE VALUE INVESTOR
The 8-Puzzle
To apply these ideas to AI, we must first medi-
tate on the activities of a value investor. I focus analysis
on the discretionary investor who holds a concentrated
portfolio over long horizons in the style of Benjamin
Graham or Warren Buffett. This investor does not trade
frequently and seeks to purchase undervalued securities
at a margin of safety that will deliver abnormal returns
over long horizons. This analysis may also apply to
discretionary growth investors, but it will not fit the
short-term traders who trade in and out of stocks at
high volume. I focus on the equity markets as a case
study. All of the analysis in this article may apply to the
empty space (Slocum and Sonneveld [2006]). One way discretionary bond investor.2
to represent this puzzle is to represent the location of Consider the value investor. This investor seeks
each tile, but an even better representation is to repre- to purchase securities at favorable prices, hold them
sent the location of the hole. The representation that for long periods, and sell them after appreciation. The
the analyst employs will determine which algorithm portfolio of securities is small (e.g., fewer than 20), and
to write. much of the work goes into selecting the right security
The core of AI is the agent, which is similar to the at the right price. We can represent the state as a vector
rational agent that populates models of academic eco- (s, t, p, b, e), where s is the security, t is the time, p is the
nomics, such as in game theory and general equilibrium current market price, b is the purchase (buy) price, and
(Osborne and Rubinstein [1994]; Mas-Colell, Whinston, e is the sales (exit) price. For value investors like Warren
and Green [1995]). The agent behaves according to Buffett, this exit price does not need to be defined. 3
the actions available and takes the state of the world The action of the investor at any given state is to buy,
as an input. The agent can transition to new states and sell, or hold. Buy and sell refer to transactions on the
obtain new information on its environment (or even security, and hold refers to doing nothing. Because stock
develop new actions). Much of successful AI is effi- prices constantly change, the state changes even when
ciently representing the problem and the agent’s strate- the investor takes no action.
gies, actions, and preferences. To visualize this, consider the decision tree in
The first problem of the artificial agent is defining Exhibit 2, where each node pertains to a state of the
the state space. The state space captures a representation of world. For the sake of exposition, I will consider only
the world as it pertains to the problem at hand. The next the case in which the investor seeks to purchase one
step is to define the action set of the agent. This is how the stock. For example, the state in Exhibit 2 could be
agent moves between each state. Finally, the analyst must (Apple, March 25, 2016 at 2:00 p.m., $90, $80, $150).
define a goal test, conditions that solve the problem. The The investor’s choices are to buy Apple, sell Apple if he
transition model determines how actions move the agent or she already owns it, or hold it and do nothing. Each
into a different state. For example, in the game of chess, of these actions leads to a new state that refers to Apple’s
the state space is the position of the pieces at any given price at 2:15 p.m. on the same day.
time. The action spaces are the legal moves available to A goal state for the investor would be a sufficiently
any player given the current board state. The transition e −b
model constitutes the rules of the game. high return on capital. Define return as R = , the
b
For our purposes, the three domains that are most excess of exit price over purchase price as a propor-
useful in AI will be searching, learning, and reasoning tion to the purchase price. An investor with a threshold
(Russell and Norvig [2010]). These are the three main return of R* will take action when the return on that
(although not only) categories that are leading research stock exceeds the threshold (R > R*). In that case, the
in AI (Segaran [2007]). investor has achieved his or her goal and can move on
to another opportunity to deploy capital.

22    A rtificial Intelligence and Value Investing Spring 2018

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Exhibit 2 up to 35 possible moves from each side. Therefore, chess
The State Space and Tree has a branching factor of 35: The tree can expand into
35 possible ways at any given node.5 Each move itself
creates a different path of the game. After any possible
move, there is a subtree from that point onward.
Exploring each subtree is infeasible even with cur-
rent computing technology. As such, researchers have
developed heuristics that approximate the value of these
subtrees (Newell and Ernst [1965]; Mitchell [1982]).
Some moves are probabilistically superior to others
because they are more likely to end in wins than losses.
A heuristic would provide some rules of thumb that
would assist in pruning the tree by eliminating subop-
timal moves. Thus, sacrificing a queen with no strategic
or material benefit would count as a simple heuristic.
There are many possible such heuristics, and much of
the details on how computers win against world cham-
This is the beginning of a model that can provide pions in chess rely on the implementation of heuristics
some high-level architecture for designing an AI- and how the machine accesses prior databases of games.6
inspired approach to value investing. The details of this Value investors also use heuristics. A value investor
point rest on implementation. We now turn to alterna- seeks to buy a stock at a discount when its intrinsic value
tive approaches within AI given this high-level scheme. exceeds its current market price. Evaluating the goal test
for an investor requires the passage of time because the
truth behind the stock is revealed years later when the
SEARCH
investor sells the stock at either a gain or a loss. However,
Representing the value investor as a tree leads to the investor must decide now whether to buy and cannot
the problem of search.4 Much of the early work in AI lies walk down the game tree into the future to make that
within search algorithms on trees, as in Exhibit 2. These decision. A heuristic that most value investors agree on
trees can be finite or infinite. The most common algo- is simple financial multiples or ratios.
rithms provide different ways of navigating a game tree The two most common ratios are price to earn-
to locate a solution that satisfies the goal state. Common ings and price to book, which are computed from
solutions are breadth-first search and depth-first search stock market and accounting data. These are heuris-
(Dijkstra [1959]). More-advanced algorithms provide tics because they are not perfect measures of whether a
heuristics that simplify the search problem (Dechter and stock is undervalued. Instead, they are approximations
Pearl [1985]; Hart, Nilsson, and Raphael [1968]). that provide assistance to the investor to make decisions
For value investing, the search problem has both today. Other value investors have defined their own set
quantitative and qualitative aspects. The quantitative of heuristics. For example, Benjamin Graham writes in
dimension is fairly straightforward: The value investor The Intelligent Investor that he seeks stocks that have net
cares about purchasing stocks at a price below their current assets per share below 67% of the purchase price,
intrinsic value. This is hard to know with certainty, common during periods of steep market declines. Other
so investors rely on common ratios or multiples. This disciples of the Benjamin Graham approach, like Charles
is the approach Benjamin Graham [2005] proposes in Brandes, have developed a four-point test for a value
The Intelligent Investor. These are similar to the heuristics stock that relaxes some of the Graham metrics and sup-
used in AI because they are shortcuts that help narrow plements with others (Brandes [1998]). Value investors
the search universe to find the stock that would satisfy are already using heuristics, and the first step is simple:
the goal test. to embed these heuristics inside an AI representation.
For a simple example of a heuristic, return to the As Buffett once noted, though, the quantita-
game of chess. For any given board position, there are tive analysis of a stock is straightforward relative to

Spring 2018 The Journal of Investing    23

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 the qualitative analysis (Ferraro [2009]). Inspection of manager or the firm. These data, up until now, have
Buffett’s stock selection over his lifetime reveals that been relatively structured in that they take a numerical
many of his selections would not qualify as deep value and are easily readable by a machine. But the undiscov-
stocks but, rather, have features of growth stocks. For ered country lies in unstructured data: namely, natural
example, when Buffett bought Disney and Coca Cola in language, social media, and data on human productivity
1966 and 1988, respectively, both were enjoying price- and output. In fact, 80% of business-relevant informa-
to-book ratios well above 1.0, the classic threshold for tion originates in unstructured form (Grimes [2008]).
a value investor. A heuristic that incorporates qualita- Indeed, the analysis of unstructured data is where AI
tive metrics must use more than the existing heuristics methods can excel because it requires the facility of
employed by value investors. This is where AI can shine. computers to process huge amounts of data that are
Observing market price is easy, but measuring unknowable by any single human (Marsland [2009]).
intrinsic value is the challenge. To make this qualita- As a simple taste of what the future will bring,
tive assessment precise, the investor needs to define the observe the dramatic improvements in sensor tech-
features, or qualities, of the company that have high nology, which have exploded the volume of data coming
intrinsic value. These can include characteristics such from wireless sensors. For example, new cars today feed
as the business model, competitive advantage, strong huge amounts of information on the performance of the
management, or good governance (Ashworth [2016]). vehicle to enable features like adaptive cruise control
In the first step, the investor needs to describe these in (Thrun et al. [2006]). These sensors will only increase
verbal terms. For example, suppose that high-value com- in scope and decrease in cost, thereby increasing the
panies have managers with long horizons, large owner- amount of data available for analysis. The Internet of
ship stakes, deep industry experience, and arm’s length Things revolution will catapult cheap sensors into all
relationships with their boards of directors. facets of everyday life so that human productivity is
The second step is for the investor to define and measurable and quantifiable. These changes can already
identify quantitative measures of these attributes. In be seen in the way a Fitbit or Nike Fuelband can quantify
my example, they would be the length of the vesting aspects of a person’s fitness or health (Watson [2014]).
cycle of stock options or restricted stock, the ownership As an example, consider a sales manager at a com-
stake in the company measured through proportionate pany. Today, the manager’s compensation depends on his
share from stock option holdings, the manager’s years of or her sales calculated on an annually or quarterly basis,
experience in the industry in which the firm operates, but these are output measures only. Once human pro-
and the number of independent directors on the board. ductivity becomes cheaper to measure, input measures
These measures are all available from public datasets. can also be incorporated into compensation. Imagine
Much of the academic accounting and finance litera- measuring his or her productivity by the number of
ture seeks to correlate these attributes with long-term emails sent to follow up on leads or the number of meet-
abnormal stock performance. ings or phone calls with prospective clients. Some com-
The next step would be to increase the size of the panies, such as Shuttle Express, Inc., have already began
input data into the AI system. To continue my example, to increase the rate at which they monitor employees
suppose managers who give conservative forecasts or dis- (Katz [2015]). In the long term, third-party compa-
cuss their firm in conservative language are more likely nies will distribute highly aggregated metrics on firms
to represent a value company than a growth company. As or industries with respect to the productivity of their
such, reading both the level of the earnings forecast from employees. The astute value investor could then use
management guidance during the conference calls and these data to help assess the quality of the company’s sales
the text from the conference call itself could provide a force and provide a better estimate of intrinsic value.
qualitative assessment of whether the company qualifies as These ideas are speculative because they are based
a value stock. Again, this area has already begun to appear on data that do not yet exist. If current trends are any
in academic research and is one of the most active areas indication, though, the measurement of economic
of research today, called textual mining (Belsky [2012]). output will only increase, thus increasing the amount
The f inal step would be to include an ever- of available data. As technology marches on, the cost
expanding amount of data into the valuation of the of acquiring new measures of corporate and human

24    A rtificial Intelligence and Value Investing Spring 2018

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 performance will plummet, and it will be imperative for picks and the universe from which he or she selected
successful investors to assemble, aggregate, process, and these stocks). Each stock could be classified as buy, hold,
make decisions over this ever-expanding trove of data. or sell. Once the analyst has trained the machine over
In this sense, the search problem of the value investor this training set, the machine would rely on its initial
will be the core problem, and advances in AI technology training to classify future stocks.
on search techniques will assist the value investor in One problem with this approach is that it may
better measuring intrinsic value. suffer from insufficient data. If the only input is the
The search algorithms discussed in this section universe of stocks and the only output is the actual
were based on human guidance. These are the handful stock selected, there is a wide chasm between input and
of metrics that successful value investors have used to output and much can be lost. The machine will have
access quality in companies. The promise of AI is to difficulty crossing the chasm, especially because training
develop new heuristics and methods that no human sets will be small; most value investors hold concentrated
could ever devise. To do so, the machine must learn portfolios over a long horizon, so the actual output set
from the data, and this brings us to our next key intel- is small. To improve training and machine learning, the
lectual pillar of AI. analyst should supply more intermediate data that are
used in the decision-making process of the value investor.
LEARNING For example, most investors employ something of
a pyramid-like decision process, narrowing the universe
The second main cluster of approaches in AI are of stocks over time. For example, quantitative filters
related to learning (Mitchell [1997]). Machine learning may sort the full universe of stocks down to a cluster of
has exploded as a field in the last decade and has many potential investments. The investor may further narrow
applications to financial markets. For example, hedge this cluster based on his or her own expertise or interest.
funds like Two Sigma Investments and Renaissance At the penultimate stage, the investor does a deep dive
Technologies are reputed to have used machine learning into the company’s financials to research one or two
techniques to screen and select diverse portfolios of investments that are candidates for a major position.
stocks for years (Vardi [2016]). How can this body of If the analyst can supply data about this sequential deci-
knowledge be useful to concentrated value investing? sion process over time, the machine could use this richer
It is important to first consider the main forms of data structure to improve on its own learning.
learning. The two broad categories are supervised and An open question is whether the machine would
unsupervised learning. need to represent the sequential narrowing process
explicitly. This is where the field of AI has differing
Supervised Learning opinions. There are those who believe all that matters
is the final output, regardless of the specific approach
The classical and original form of machine learning or method used. Those in this group, often called
took place under human supervision. A human classifies scruffies, are willing to use any manner of statistical or
a set of input–output pairs according to a hypothesis. mathematical tools, even if they have little resemblance
Once the machine has learned based on the training to the actual decision process at hand.8
set, it then proceeds to classify untrained data based on Another approach is to represent the decision pro-
its initial training. cess explicitly through data structures and algorithms
Consider visual recognition from photographs. that bear a faithful similarity to the actual human
Imagine 10,000 photographs of the jungle, and a process. Adherents of this view, often called neats for
human has identified 100 photos that contain a tiger. their taste in neat analytical models, draw on explicit
The training set is the 100 photos that the computer mathematical and logical axioms and work from first
knows contain a tiger. The computer then searches the principles aiming to understand key truths through any
remaining 9,900 photos and tries to classify them as given problem.
{tiger, no tiger}.7 To apply this to value investing, a nec- Both approaches are needed, and the best algorithm
essary input would be a relevant training set (i.e., data on will benefit from a combination of both approaches.
a successful value investor and both the investor’s stock A more structured representation would take each round

Spring 2018 The Journal of Investing    25

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 of the decision process as an input into some explicit data problem is to find stocks of high intrinsic value and low
structure, and the machine would learn from its datasets price. Estimating intrinsic value requires the investor to
how to refine every step. This of course requires larger screen large amounts of information, both qualitative
inputs, and it would require an investor to hand over not and quantitative, to deliver an assessment.
just final stock picks but even intermediate picks. It is A reinforcement learning approach would not just
likely that every value investor goes through something simulate specific selections of stocks, but rather would
of a sequential elimination process just described, and simulate different attributes that may or may not lead to
each could furnish such details if he or she had sufficient a better assessment of intrinsic value. For example, rather
financial incentive. than relying on a structured model that says long-term
compensation and independent directors lead to better
Unsupervised Learning performance, the machine would traverse the universe
of corporate attributes to discover which ones lead to
A more recent approach to machine learning skips high abnormal returns.
the initial step of human supervision, and the com- In this sense, unsupervised learning pertains not
puter learns on its own (Bishop [2007]). This requires to the selection of the securities per se, but rather to the
even larger datasets because there is no structure in the selection of the attributes that then lead to the selection
problem at all. The machine tries different approaches of the securities.9
and receives feedback from the system regarding which
ones work and which ones fail. This is called reinforce- REASONING
ment learning because the system itself provides reinforce-
ment feedback to the machine. For example, a machine Approaches thus far may seem unconventional,
learning to play chess would try different moves, and but they have already gained traction in industry. The
with enough simulations of the game, it would receive truly revolutionary work in AI concerns reasoning
feedback on which moves probabilistically are more (Minker [2000]). Getting a computer to reason logi-
profitable. Rather than relying on a human to inform cally has been a goal of computer science for decades.
the computer that losing a queen is a poor move, the This approach takes a different path than the searching
computer simulates games in which it loses its queens and learning algorithms discussed. Now, a computer and
and learns that, statistically, these games are harder to machine seek to make logical statements and inference
win. For such a strategy to work with value investing, using the same principles of logic that underlie phi-
the computer would be required to simulate stock selec- losophy and abstract mathematics (Davis [1990, 2005]).
tion and would receive market feedback on those stocks. Consider the following examples that can illustrate
This is more challenging than the chess example because elemental reasoning. A hunter seeks to locate a tiger
chess is deterministic (nonrandom), whereas the stock hiding in a cave.10 The man cannot see in the dark cave
market has uncertainty and risk at every point. A com- but can smell the tiger from a distance. He has a rif le
puter can simulate different moves in chess and, there- with one bullet and seeks to minimize his travel inside
fore, populate an entire game tree; it cannot do so with the cave so that he kills the tiger. This can be represented
the stock market without making strong assumptions as a 3 × 3 grid, as shown in Exhibit 3. Each square in
that themselves may violate the fundamental random- the grid corresponds to a location in the cave, and the
ness of markets. man starts at Square 1. He can move up or down, but
As such, an alternative strategy is to mine past data not diagonally, one square at a time. The tiger sits in
using historical stock prices as the ultimate output. To do Square 6, but the man cannot see the tiger. Instead, he
this, the investor would need to back test the algorithm can smell the tiger in any adjacent square (again, up or
on the last 50 years of stock data. This approach is pos- down, but not diagonal). From the starting position, the
sible, although I do not believe it to be of high value man smells nothing, so he knows for sure that the tiger is
given the long horizons and low turnover of successful not in Squares 2 or 4. Therefore, he moves into Square 2
value investors. with certainty that he will not be eaten. Again, he smells
The third approach is to use unsupervised learning nothing and infers that the tiger is not in Squares 1, 3,
to improve search heuristics. Recall that the search or 5. If it were, he would have smelled the tiger while

26    A rtificial Intelligence and Value Investing Spring 2018

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Exhibit 3 (Hwang and Schubert [1993]). The example given indi-
A Reasoning Game cates a level of reasoning that is common in everyday
human life but still foreign to machines. Investing in a
stock market involves layers of uncertainty and infor-
mation that is asymmetric. Yet, there are still parallels.
A computer could begin to reason about managerial
attributes and corporate quality based on what is and
what is not disclosed in its financials. For example,
during an economic downturn, one would expect cut-
backs in research and development expenses. If no such
cutbacks are forthcoming, this may reveal that the firm
is more insulated from macroeconomic events, leading
to a higher intrinsic value during times of distress. This
is an example of how reasoning and inference can be
employed through machines to draw conclusions about
a company’s quality.

LOOKING FORWARD

Logic-based AI may still be a few years away,

standing in Square 2. Of course, he already knows there but there is no doubt that it is in our future. However,
is no tiger in Square 1 because he started there. making this future a reality will require several steps
Now, suppose he moves down to Square 5. Because along parallel tracks. These steps need not be coordi-
the tiger is in Square 6, he smells the tiger. However, nated or happen at the same time, but they do need
smelling a tiger in Square 5 means the tiger could be in to happen. First, the value investing community will
any of Squares 2, 4, 6, or 8. Of course, the tiger is not need to acknowledge that the application of AI to their
in Square 2 (because he was just there), but it is also not field is inevitable and embrace rather than fight this
in Square 4. If it were in Square 4, he would have smelled secular trend. Rather than see this as a war between man
it in Square 1. Therefore, the tiger must be in Squares and machine, the discretionary investing community
6 or 8. Suppose he moves left to Square 4. He smells needs to acknowledge that computation can assist in the
nothing because there is no tiger in Squares 1, 5, or kind of reasoning traditionally done by humans, thus
7. When he moves down to Square 7, he again smells enlarging the scope of analysis. Investors will need to
nothing and concludes that the tiger must not be in adjust their analysis, knowing that they will now have
Square 8. If it were in Square 8, he would have smelled a very powerful tool, an AI engine, to use along with
it in Square 7. Therefore, by the process of elimination, their own minds.
the tiger must be in Square 6 because he smelled it in First, investors need to think hard about the mental
Square 5 but not in Square 7. The man returns to Square process through which they make their investments.
5 and shoots the tiger in Square 6.11 Successful investors have already done this, as many of
This simple example from a child’s puzzle illus- the best investors have written books and given speeches
trates some of the intuitions and layers of logic that a about their investment process. Others have spoken at
computer could conceive, construct, and execute. The length about specific investments, which reveals this
problem is well defined; therefore, building a data struc- investment process to outsiders. This first step must come
ture and algorithm is straightforward for such a setting. from the investing community if AI is to have a legiti-
As with much of AI, however, this simple example can mate application to value investing. Investors will need
illustrate some techniques that may apply more broadly to meditate on the kinds of data sources they bring to
in real-life settings. bear on their investment choices and how exactly they
The core idea is to use information to reason, interpret those data. Investing is ultimately a process of
to take actions, and to infer the state of the world reducing down an enormous set of financial information

Spring 2018 The Journal of Investing    27

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 into qualitative statements about price and quality. and loss), then the computer can essentially back-solve
If a few value investors were to write down in narrative and improve itself over time.
form how this process takes place, that could provide the This process is similar to how machine learning
initial material for computer science to step in and build has progressed. In the early phases, machine learning
a more specific AI model tailored to value investing. algorithms drew upon a training set based on human
On the technology side, computer science as a classification; in later, more mature versions of machine
field will need to continue its steady march toward learning, algorithms were able to learn without a com-
logic-based AI. This shows no signs of abating, but the prehensive training set. In those latter cases (reinforce-
research is still in its early stages. Although existing ment learning), the computer would experiment and
methods of AI (such as machine learning) have leaned simulate different actions while improving performance
heavily on statistics, logic-based AI will lean more (in a precise, statistical sense). Indeed, the success of AI
heavily on pure mathematics and philosophical logic. in games like chess and Go is a shining example of this
This is an entirely different branch of knowledge and process. Once these two parallel developments occur,
currently will require greater inroads into cross-disci- they will propel both AI and value investing forward.
plinary research. Beyond this, computer scientists will Some of these steps forward will come from university
need to communicate extensively with value investors researchers, and others will come from the investing
to build a custom AI engine for investing. They will and technology communities. Today, these communi-
need to take the narrative descriptions documented by ties have been largely separated, and much of the cur-
the value investors and program their engine to replicate rent applications of AI to investing come from within
this kind of reasoning on a large scale. Although this finance, rather than from within technology. With the
may seem daunting, the Wumpus World is an example next iteration, I believe there needs to be a closer align-
of one successful case. In that children’s game, the rules ment between the two fields, with innovations coming
provide enough structure for a computer algorithm to as much from technology as from finance. Regardless,
work through. A similar but more elaborate process will I see this future as inevitable and am hopeful for both
need to occur for an investing example. communities. This process will push technology for-
Investors who have documented their data sources ward to make markets more efficient. As always, the
for a variety of successful investments will serve as the early movers in this space will reap the highest returns.
guiding light for the AI engine. This corpus of knowl-
edge will be hardcoded into an AI algorithm, and the CONCLUSION
algorithm will essentially replicate the human reasoning
process through formal rules and explicit procedures At the end of the day, value investing is about
written in software. As the engine applies this tech- rationality in the face of irrational sentiment and mass
nique to actual investing, early results will be mixed, investor psychology (Brandes [1998]). It is precisely the
and there will be many misclassification errors (suc- irrationality of human investors that drives prices away
cessful investments incorrectly classified as poor, and from fundamental value. Therefore, computers that can
poor investments classified as successful). However, as reason ever more faithfully than any human would have
data improve and as more reasoning methods are hard- a natural advantage in such an environment. These ideas
coded into software, results will also improve. of machine reasoning are in their earliest stage of devel-
Later, the AI engine will learn to reason on its opment but will have the most market impact on value
own in ways that do not easily replicate what humans investing in the long term.
do. After incorporating much of the early intuitions and The strong form of the efficient markets hypoth-
simple logic of human-based reasoning, the software esis states that market prices will impound all available
will obtain a level of complexity that no human mind information. Even the most fervent supporters of efficient
can fathom and will make inferences and decisions in markets do not believe that markets today are strongly
ways that are unknowable (and possibly incomprehen- efficient. Nonetheless, the continual advancement of tech-
sible) to human investors. If the AI engine successfully nology in computation brings more data to bear on the
establishes its goal tests and a clear metric of success investing problem. Sophisticated investors can make use
(relatively easy in the black and white world of profit of this information that is not yet impounded into price.

28    A rtificial Intelligence and Value Investing Spring 2018

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 8
There are enormous gains from moving forward in Google’s algorithm recently beat the highest-ranking
this direction. It is smart investors who make the market professional Go player. The research behind the specific
efficient; they identify opportunities where value and methods used in Google’s algorithms show that it was an
price differ, and their trades close this gap. Artificial alphabet soup of assorted AI techniques pooled together that
intelligence is the primary theoretical apparatus neces- dominated the human opponent. This is an example of a
scruffy approach.
sary for processing the ever-exploding amounts of data 9
Critics of this approach call this data mining because it
emerging from human and corporate behavior. Making lacks theory and instead runs large-scale statistical analysis to
sense of all these data, and reasoning through them, will determine economic outcomes.
remain the biggest challenge for 21st-century investing. 10
This is a variation of the Wumpus World Puzzle in
In this article, I hoped to capture some of the spirit of elementary logic.
this challenge and to outline the existing areas within 11
Of course, this is not the shortest path. Had he moved
AI that could be profitably deployed for a value investor. to Square 3 from Square 2, he would have smelled the tiger
This is just the beginning. from Square 3 and identified it in Square 6, but that occurs
because Square 3 is on the edge.
ENDNOTES
REFERENCES
I’d like to thank Brian Bruce and Dylan Shell for inspi-
ration and comments. Ashworth, W. “Qualitative Analysis: What Makes a Com-
1
I refer to computers and technology equivalently pany Great?” Investopedia, July 11, 2016. https://www
as machines. Computer remains an outdated term because .investopedia.com/articles/fundamental-analysis/10/
machines today take the form of smartphones, cloud services, qualitative-analysis-intangibles-affect-stock-price.asp.
desktop computers, mainframes, and so forth.
2
However, bonds themselves have a richer mathematical Belsky, G. “Why Text Mining May Be the Next Big Thing.”
structure than equities and lend themselves to quantitative Time, March 20, 2012. http://business.time.com/2012/03/20/
analysis that permeates existing algorithmic trading. I focus why-text-mining-may-be-the-next-big-thing/.
on equities because of their more qualitative component rela-
tive to bonds. Of course, these are arbitrary distinctions, and Bishop, C.M. Pattern Recognition and Machine Learning.
a worthwhile agenda would be to apply AI to bond investing Springer-Verlag, 2007.
as well.
3
This is part of what made Warren Buffett so successful. Brandes, C.H. Value Investing Today, 2nd ed. New York:
He stated that, “If you instead focus on the prospective price McGraw Hill, 1998.
change of a contemplated purchase, you are speculating”
(Farrell [2014]). Davis, E. Representations of Commonsense Knowledge. San
4
The tree is an example of a directed acyclic graph. It Mateo, CA: Morgan Kaufmann, 1990.
is a graph because it contains nodes and links between nodes.
It is directed because the links can move in one direction as ——. “Knowledge and Communication: A First-Order
time moves forward, and it is acyclic because there are no Theory.” Artificial Intelligence, Vol. 166, No. 1–2 (2005),
cycles. Nodes cannot link back to each other because time pp. 81-139.
only moves forward.
5
Although large, this is still smaller than the game of Dechter, R., and J. Pearl. “Generalized Best-First Search
Go, which has a branching factor of 270. Strategies and the Optimality of A*.” Journal of the Association
6
This technology was used by IBM to create Watson, of Computing Machinery, Vol. 32, No. 3 (1985), pp. 505-536.
the supercomputer that competed on the Jeopardy game show.
See Ferrucci et al. [2010]. Dijkstra, E.W. “A Note on Two Problems in Connexion with
7
This is how the iPhoto software on the Apple Macin- Graphs.” Numerische Mathematik, 1 (1959), pp. 269-271.
tosh works. The program asks you to classify photos and iden-
tify people by name in a handful of photos, and the computer Farrell, M. “Four Long-Term Investing Tips from Warren
then learns from this training and seeks to classify by itself on Buffett.” The Wall Street Journal, February 28, 2014.
the larger untrained dataset.

Spring 2018 The Journal of Investing    29

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Ferraro, S.R. “The Buffet Approach to Valuing Stocks.” Mitchell, T.M. “Generalization as Search.” Artificial Intelli-
Graziadio Business Review, Vol. 12, No. 3 (2009). gence, Vol. 18, No. 2 (1982), pp. 203-226.

Ferrucci, D., E. Brown, J. Chu-Carroll, J. Fan, D. Gondek, ——. Machine Learning. New York: McGraw-Hill, 1997.
A.A. Kalyanpur, A. Lally, J.W. Murdock, E. Nyberg, J. Prager,
N. Schlaefer, and C. Welty. “Building Watson: An Overview Newell, A., and G. Ernst. “The Search for Generality.”
of the DeepQA Project.” AI Magazine, Vol. 31, No. 3 (2010), In Information Processing 1965: Proceedings of the IFIP Congress 65.
pp. 59-79. Washington, D.C.: Spartan Books, 1965.

Graham, B. The Intelligent Investor: The Classic Text on Value Osborne, M.J., and A. Rubinstein. A Course in Game Theory.
Investing. HarperBusiness, 2005. Cambridge, MA: MIT Press, 1994.

Grimes, S. “Unstructured Data and the 80 Percent Russell, S., and P. Norvig. Artificial Intelligence: A Modern
Rule.” Breakthrough Analysis, August 1, 2008. https:// Approach, 3rd ed. Upper Saddle River, NJ: Prentice Hall,
breakthroughanalysis.com/2008/08/01/unstructured-data- 2010.
and-the-80-percent-rule/.
Segaran, T. Programming Collective Intelligence: Building Smart
Hart, P.E., N.J. Nilsson, and B. Raphael. “A Formal Basis Web 2.0 Applications. Sebastopol, CA: O’Reilly, 2007.
for the Heuristic Determination of Minimum Cost Paths.”
IEEE Transactions on Systems Science and Cybernetics, Vol. 4, Slocum, J., and D. Sonneveld. The 15 Puzzle. Slocum Puzzle
No. 2 (1968), pp. 100-107. Foundation, 2006.

Hwang, C.H., and L.K. Schubert. “EL: A Formal, yet Thrun, S., M. Montemerio, H. Dahlkamp, D. Stavens,
Natural, Comprehensive Knowledge Representation.” Paper A. Aron, J. Diebel, P. Fong, J. Gale, M. Halpenny,
presented at symposium, AAAI, 11th National Conference on G. Hoffman, K. Lau, C. Oakley, M. Palatucci, V. Pratt,
Artificial Intelligence, Washington, D.C., July 11–15, 1993. P. Stang, S. Strohband, C. Dupont, L.E. Jendrossek,
C. Loelen, C. Markey, C. Rummel, J. van Niekerk, E. Jensen,
Katz, M.L. “Monitoring Employee Productivity: Proceed P. Alessandrini, G. Bradski, B. Davies, S. Ettinger, A. Kaehler,
with Caution.” Society for Human Resource Management, A. Nefian, and P. Mahoney. “Stanley, the Robot that Won
June 1, 2015. https://www.shrm.org/hr-today/news/hr- the DARPA Grand Challenge.” Journal of Field Robotics,
magazine/Pages/0615-employee-monitoring.aspx. Vol. 23, No. 9 (2006), pp. 661-692.

Lin, S. “Computer Solutions of Travelling Salesman Vardi, N. “Top Quant Hedge Funds Stand Out with Good
Problem.” Bell Systems Technical Journal, Vol. 44, No. 10 2015.” Forbes, January 13, 2016. https://www.forbes.com/
(1965), pp. 2245-2269. sites/nathanvardi/2016/01/13/top-quant-hedge-funds-
stand-out-with-good-2015/#513b6b75c602.
Marsland, S. Machine Learning: An Algorithmic Perspective,
1st ed. Boca Raton, FL: CRC Press, 2009. Watson, C. “The New New Thing: The Internet of Things.”
OZY, December 26, 2014. http://www.ozy.com/fast-forward/
Mas-Colell, A., M.D. Whinston, and J.R. Green. Microeco- the-new-new-thing-the-internet-of-things/32714.
nomic Theory. New York: Oxford University Press, 1995.

Minker, J. Logic-Based Artificial Intelligence. Boston: Kluwer, To order reprints of this article, please contact David Rowe at
2000. drowe@ iijournals.com or 212-224-3045.

30    A rtificial Intelligence and Value Investing Spring 2018

Downloaded from https://pm-research.com/content/iijinvest/27/1, by guest on March 15, 2024. Copyright 2018 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.