1.
Q-Group Fall 2017 Monday Oct 16 Special Track on Machine Learning
Presenter Walter Alden Tackett, CFA

2.
Introduction: your time is valuable
• You’re covered for prerequisites of Machine Learning
You are a specialist in quantitative finance
• Mastery requires time, effort, and opportunity cost
Any tool for investment must first be mastered
• There are mixed signals with extreme values
Is Machine Learning worth your time?
• The differences are part of a bigger picture.
ML and Quant Finance have the same lineage.

3.
#DEFINE
A Key to Acronyms and Abbreviations
Machine
Learning
Computer
Science
OR
Operations
Research
EE
Electrical
Engineering
CS
Artificial
Intelligence
nn
Neural
Networks
IT
Information
Technology
BS
Big Data
Science
CS EEOR
AI NN IT BS
Acronyms Only: terms will be defined where needed
ML

4.
About the
term
Machine
Learning
Priority: Used by Arthur Samuel in
the title of the first journal article
describing the first ML program.
•Called Statistical Learning by
… (spoiler alert) … statisticians.
Related, not
synonymous:
Statistical Data
Analysis (SDA,
due to Tukey)
and recently
Data Science are
more inclusive:
•ML, Very-Large
Databases, and
Visualization
Graphics all fall
under SDA and
Data Science.
Problem: The word Learning is
anthropomorphic …
•… which contributes to the AI hype-
cycle.
Adaptive
Sampled-Data
Statistical
Computing
would be most
descriptive.
•… but not many
people would talk
to you.

5.
Machine Learning, Big Data, Data Science,
and Artificial Intelligence: Hard questions
ML successfully
recommends
Amazon & Netflix
products – how is
that relevant to
price returns?
Algorithms can
learn if email is
spam or not
spam – does that
translate to buy /
sell / hold?
Does ML always
require large
datasets, labor-
intensive labeled
samples, or
categorical data?
Where are the
Quant Finance
“Wins?”

6.
Machine Learning, Big Data, Data Science,
and Artificial Intelligence: Maybe not?
Isn’t there some kind of
sketchy relationship
between Machine
Learning and:
•Data Mining?
•Back-testing?
•Expert Systems?
1
Do market data’s well-
known issues thwart
“Big Data Science?”
•What qualifies data as Big
data?
•Has all useful information
already been strip-mined
from financial data?
2
Isn’t this maybe the 3rd
generation of AI-
related investment
hype?
• Neural Networks… again?
Only “deeper?”
• A bigger box of logistic
regression and fairy dust?
3
Phoenix Wright: Ace Attorney Copyright © 2001-2017 Capcom Co., Ltd.

7.
Machine Learning and Artificial Intelligence:
Reviewing the Fear, Uncertainty, and Doubt
Have you heard the
news?
• “Goldman Sachs
replaced 100 traders
with 40 computer
engineers. Or 20, or
something.”
• Bonus points: “There
will be no jobs for
humans by 20xx.”
AI is a threat to
humanity
according to…
• Elon Musk
• Stephen Hawking
• Nick Bostrum
• All books on the
Amazon AI Threats
to Humanity list*
Should Quant
Investors be
concerned about…
• Disruption similar to
effect of HFT on
trading?
• Computers beating
human champions at
poker?
• AI taking their jobs?
• Human extinction?
* List does not currently exist on Amazon, but consider that an AI program may be deciding what topics get their own list.

8.
BS Bonanza: commodity Big Data Science is
not a starting point for investment strategy
Problem:
Businesses
have large
stores of
data to mine,
much of it
unstructured
Innovation:
Powerful,
free, easy-to-
use
statistical
packages are
available -
just add data
Demand:
Workers are
needed who
can run the
packages on
the data,
ASAP
Result:
Big Data
marketed,
sold, bought,
implemented
on minimum
time & cost
basis
Precedent:
US
Businesses
converted CS
and
computing to
commodity
minimum-
cost IT model
in the 2000’s
Effect:
A new Big
Data Science
(BS) model
compounding
ignorance of
CS, statistics,
and bona fide
data science
Conclusion:
Like the IT
model, the
BS model and
associated
hype aren’t
predictive of
investment
success
Prescription:
Define Machine Learning from first principles, and relate
it to Quant Finance practice

9.
First Principles

10.
Machine Learning systems optimize
their response to input using input
samples and automated adjustment.
The Policy maps input to response
based on its Parameters. The Loss
Function measures map quality.
Update of Policy Parameters
minimizes Loss for expected out-of-
sample data …
Optimizing for out-of-sample expectation is called generalization.
Generalization is the central problem of ML :
• How and why will unobservable data differ from in-sample data?
See: First page of the first article about the first ML program*.
* Samuel, Arthur L. “Some Studies in Machine Learning Using the Game of Checkers.” IBM Journal of Research and Development 3, no. 3 (1959): 210–229.
… “from the future”

11.
Generalization: policy updates that
minimize expected out-of-sample loss.
… which must generalize from in-sample data.
Out-of-sample data is unobservable on update …

12.
Generalization: policy updates that
minimize expected out-of-sample loss.
Infer bird B∈{f, w, s} based on examples
Ask a Human: write a program to solve White Swan.
… which must generalize from in-sample data.
Out-of-sample data is unobservable on update …

13.
Generalization: policy updates that
minimize expected out-of-sample loss.
Infer bird B∈{f, w, s} based on examples
Ask a Human: write a program to solve White Swan.
Multiple predictors: color & height …
• Idea: Use “many” features. Are some better than others?
Rank: not just on in-sample performance
• Consider: shape properties invariant to rotation, scale, noise ...
… which must generalize from in-sample data.
Out-of-sample data is unobservable on update …

14.
Generalization: policy updates that
minimize expected out-of-sample loss.
Infer bird B∈{f, w, s} based on examples
Ask a Human: write a program to solve White Swan.
Multiple predictors: color & height …
• Idea: Use “many” features. Are some better than others?
Rank: not just on in-sample performance
• Consider: shape properties invariant to rotation, scale, noise ...
Robust features: e2/e1, convex hull/perimeter.
• We humans “solved it.” Computers generalizing in this way is the goal of ML.
… which must generalize from in-sample data.
Out-of-sample data is unobservable on update …

15.
Use ML method for White Swan, assuming
good features and policies are unknown.
Standardize in- and out-of-sample data:
• Resize to 4096 colored pixels; Set each pixel value to 1 or 0.

16.
Use ML method for White Swan, assuming
good features and policies are unknown.
Standardize in- and out-of-sample data:
• Resize to 4096 colored pixels; Set each pixel value to 1 or 0.
Add perturbations to in-sample data.
• Adversary ML: makes data, loss func= -1*(loss of learner).
Train a system to
tag images by logo.
•Logos above from
template png files.
•Synthesize data based
on templates and
source description.
Images
forthcoming
•From public sightings
varied lighting and
perspective.
•Shot with mobile.
•Logo colors may differ.
Expect background
(mostly) removed.
•Remaining foreground
pixels number between
N/4 and 4N.
•N based on foreground
pixels in templates.
Use-case for template-based expected-distortion sampling
or adversarial data generation:

17.
Use ML method for White Swan, assuming
good features and policies are unknown.
Standardize in- and out-of-sample data:
• Resize to 4096 colored pixels; Set each pixel value to 1 or 0.
Add perturbations to in-sample data.
• Adversary ML: makes data, loss func= -1*(loss of learner).
Data Analysis: note owl eyes, flamingo legs…
• Choose wisely: convolutional NN were literally made for 2D pattern rec.
Must separate (extended) data into at least three
partitions: train, test, and validate. Always true in ML.
Train a system to
tag images by logo.
•Logos above from
template png files.
•Synthesize data based
on templates and
source description.
Images
forthcoming
•From public sightings
varied lighting and
perspective.
•Shot with mobile.
•Logo colors may differ.
Expect background
(mostly) removed.
•Remaining foreground
pixels number between
N/4 and 4N.
•N based on foreground
pixels in templates.
Use-case for template-based expected-distortion sampling
or adversarial data generation:

18.
Train, Test, Validate

19.
Machine Learning and Regression:
Using Lasso and LARS
The White Swan example is a supervised classification problem.
• Popular misconception: ML can only be used for discrete classification
Machine Learning addresses every type of statistics used by quants
Regression & factor selection are key ML topics – a good starting point …
Regularized Regression: the Lasso and LARS
Lasso: Combines OLS with constrained optimization, minimizing MSE subject to:
sum(abs(Coefficients)) ≤ C
LARS efficiently finds unique optimal “path” of coefficients as C ranges over [0…∞)
This “budgeting” approach eliminates factors with low discriminant power
The sum(abs(…)) L1-norm tends to drive coefficients of “weak” variables to zero.

20.
Lasso in Academic Finance and
Practical use in Testing with Noise
Several recent papers apply
Lasso to factor identification.
Typically these propose Lasso
“improvements” that favor the
factor corresponding to some
economic hypothesis X.
Lasso and LARS are central to
research in sparsity (more
factors than data), with texts and
R code available for free. This
makes an accessible entry point
for quants to explore hypothesis
X hands-on.
2016 - 2017
saw interest
in the Lasso
among the
factor testing
community.
1.Create 𝟏
𝒑⁄ =100 random
time series similar to factor
fn*. Let bogus factor bn ≡ the
most significant according to
univariate regression on y.
K≥1 iterations create a family
of K bogus factors bnk~fn.
Repeat for each nÎN.
Will the real fn please stand out?
Use Lasso, LARS, and Bayes’
rule on panels drawn from
actual fn and bogus factors bnk
to test factors’ significance.
Large K supports typicality
analysis of results given 𝒑.
Lasso-based
significance
of N factor
time series
explaining y:
*Similar to factor n means random data filtered to have the same mean, variance,
higher moments, serial correlation, or other statistics matching those of factor n.

21.
Useful Things to Do:
Finding Transient Factors with Clustering
• Example of overlap between ML and Quant finance.
• Try an equity universe e.g. RU1000, N≈1000:
1. Create residual returns for each stock by removing
contributions of explanatory factors*
𝜀% 𝑡 + 1 = 𝑟% 𝑡 + 1 − , 𝑥%. 𝑡 𝑓. 𝑡 + 1
.∈2
2. Calculate empirical correlation matrix on trailing 252TD
(1-year daily) window – it’s OK that N>T
3. Perform PCA on the correlation matrix
4. Use D=2 eigenvectors eA, eB as 2D “coordinates” **
1. Note: eA and eB have one entry for each security
2. Each represents security contribution to residual correlation
* Key: 𝑡 indexes trading days; 𝑟% 𝑡 + 1 and 𝑓3 𝑡 + 1 are security and factor returns respectively for interval 𝜏 ∈ (𝑡, 𝑡 + 1], and 𝑥%. 𝑡 is exposure of security 𝑖 to factor 𝑘 at point 𝑡 in time.
Caution: Do not use residuals published in risk model flat files for this purpose: most vendors process them significantly. Also (as demonstrated) the results of varying the set K are informative.
** Clustering on D > 2 is limited in practice by significance of corresponding eigenvalues and, for K-means, the degradation of Euclidian distance measure as D grows (with typical onset in the range DÎ[7 … 10])

22.
Eigenvalues: Residual of Return Correlation
(Return ex- intercept and industry returns)
Variance
PCA of Corr as above, except with Residual
Return data replaced by samples drawn from
standard normal distribution
v1 ~ e1
v2 ~ e2
v3 ~ e3
PCA of Corr for Residual Return Data
Variance:
Principal Components of Residual Correlation vs.
Principal Components of Correlation for Noise
25
20
15
10
5
0
0 50 100 150 200 250

23.
At each point in time perform K-means
clustering in the 2D space of e1, e2
• Plot results over time for different residual types …
• Residual Return of security =
• security return ex-contribution of universe “market
intercept” (xs-mean)
• security return ex-contributions of (both):
• Market Intercept (xs-mean) of universe
• Corresponding industry return (xs-mean by industry, ex-intercept)
• random noise (e1 and e2 of corr matrix, 255 day window)

24.
Residual Return:
K-Means clustering over time
Do-it-Yourself Project: Supplementing Factor Models
Form K-means clusters on residual returns
Tabulate characteristic labels by cluster:
• Industry (and Sub-Industry – note dependence!)
• Ranked (Hi/Med/Lo) fundamentals
• Ranked (Hi/Med/Lo) technical factors
Bayesian Analysis: Are labels or their co-occurrence
persistently significantly conditioned on clusters?
random noise
security return ex-
contribution of universe
“market intercept” (xs-
mean)
security return
ex-contributions of both:
1. Market Intercept
(xs-mean) of
universe
2. Corresponding
industry return

25.
Where do we get the K in K-means?
• In K-means, the number of clusters K must be chosen
• K is a parameter: it can be overfit to in-sample data
• A separate data partition must be used to test overfit
• Results are only valid for a final partition, untouched by above process
• Idea: let K be chosen automatically for each dataset
• But choice algorithms have parameters ... it‘s turtles all the way down
• Sooner or later, parameters must be set from data
• It is a perilous undertaking, and easy to exhaust available data
• The 80‘s were the golden age of ML data snooping
• ML users faced problems long addressed by Statisticians

26.
The Modern Age of 1991

27.
Machine Learning has a long history:
1991 marks the relevant turning point
1956-
1959
Arthur Samuel’s Checkers
Program Complete, Plays
Live on TV in 1956
“Some Studies in
Machine Learning…”
Publication delayed until 1959
1961
“Since current machines
have (only 32 KB RAM), and
(maybe 1 MB RAM ten years
from now), we see that
Dynamic Programming
problems cannot be solved
(directly) because of the
memory requirements.”
• R. Bellman, Adaptive Control
Systems, 1961
Early ’60s
to Early ’80s
Symbolic AI
dominates CS
• Rule-based inference
• Symbolic reasoning
• Learning: CART
decision trees
• Main ML progress in
fields of Bio, Psych, &
EE Pattern
Recognition
1980’s-
1980-1986 – Parallel Distributed
Processing and the Neural Network
Renaissance
Neural Network models developed by
psychologists and biologists show promising
results on hard AI problems by learning from
data.
NN, and hence ML, become focus of AI
1987-
Power without
knowledge:
ubiquitous PCs
allow many NN
users to learn
about in-sample
vs out-of-sample
the hard way.
1989-
Decline of the NN
Cargo Cult /
Opening the Vaults
• Mathematical Innovation:
optimization-based
SVN and Bellman-based
Reinforcement Learning
• For problem-solvers,
pragmatism replaces
“brain-like” rationale
• EE/OR of 40’s/50’s/60’s
are mined for newly-
practical ideas
• Bio/Psych take interest
1991
“Statistical Data
Analysis in the
Computer Age”
published in
Science magazine.
This turning point to
the modern era
provides rigorous
statistical basis for ML
algorithms and
confidence testing.
1998-
Machine Learning as
Shadow Industry
PageRank statistical learning
algorithm created by Page, Brin,
Mogwani, & Winograd ca 1996
Google incorporates in 1998
Google dominates competitors
by applying ML to key problems
of Internet business 2000-2005
2016
“[Modern] neural
networks used for
machine learning ... are
generally not designed
to be realistic models of
biological function. …
[They are grounded in]
linear algebra,
probability, information
theory, and numerical
optimization.
- Goodfellow et al, Deep
Learning, 2016

28.
"Most of our familiar statistical methods, such as hypothesis testing,
linear regression, analysis of variance, and maximum likelihood estimation, were
designed to be implemented on mechanical calculators.”
"Modern ~ computers facilitate ~ new statistical methods that
require fewer distributional assumptions than their predecessors and
can be applied to more complicated estimators.”
"These methods allow the scientist to explore and describe data and
draw valid statistical inferences without the usual concerns
for mathematical tractability.”
"This is possible because traditional methods of
mathematical analysis are replaced by specially
constructed computer algorithms.”
"Mathematics
has not
disappeared
from statistical
theory.
It is the main
method for
deciding which
algorithms are
correct and
efficient tools
for automating
statistical
inference.”
Statistical data analysis in the computer age
Efron & Tibshirani, Science. Jul 26, 1991

29.
Statistical data analysis in the computer age …
coincides with popularity of nonlinear systems
Statisticians advocating
computational methods that
improve flexibility without
sacrificing statistical power…
Field: Statistical Testing and Validation
Barrier: Constraints of Mathematical Statistics
Insight: Harness the brute force of 80’s-era PCs
Innovation: Statistical Data Analysis Algorithms*
… as ML re-diversifies, NN
paradigm shifts from cargo
cult to one component of
larger toolbox informed by
CS/EE/OR.
Field: Human-competitive prediction and decision
Barrier: Limited knowledge of how brains work
Insight: NN turns to previously-impractical EE/CS
Innovation: Ideas infeasible in 40’s-60’s are doable
*Statistical Data Analysis is the term coined by John Tukey, and used by the Efron & Tibshirani for continuity; however Statistical Data Analysis Algorithms is more accurate.

30.
Statistical data analysis in the computer age …
details of the new emergent synthesis
Field: Statistical Testing and Validation
Barrier: Constraints of Mathematical Statistics
Insight: Harness the brute force of 80’s-era PCs
Innovation: Statistical Data Analysis Algorithms*
Mathematical Statistics:
Based on algebra & calculus.
Confidence in results derive
from proofs that require
analytically tractable formulas
and assumptions about data.
Statistical Data
Algorithms:
Based on brute force re-
sampling. Confidence derives
from sampling theorems,
algorithm properties, and the
statistic(s) applied to each
sample set.
Field: Human-competitive prediction and decision
Barrier: Limited knowledge of how brains work
Insight: NN turns to previously-impractical EE/CS
Innovation: Ideas infeasible in 40’s-60’s are doable
“Biological” Neural Nets:
Parallel Distributed Processing
… gets a jumpstart from
research neglected by
60’s/70’s Symbolic AI.
Practical implementations
usually based on CS, EE, OR
methods.
Golden Oldies:
Algorithms of the 50’s+90’s CPUs
Dynamic Programming, Monte
Carlo Methods, Convex
Optimization, SVM
Popularized by their use in NN,
their variants are today’s core
ML tools.
*Statistical Data Analysis is the term coined by John Tukey, and used by the Efron & Tibshirani for continuity; however Statistical Data Analysis Algorithms is more accurate.

31.
Statistical data analysis in the computer age …
specific milestone changes to the ML field
Field: Statistical Testing and Validation
Barrier: Constraints of Mathematical Statistics
Insight: Harness the brute force of 80’s-era PCs
Innovation: Statistical Data Analysis Algorithms*
Mathematical Statistics:
Based on algebra & calculus.
Confidence in results derive from
proofs that require analytically
tractable formulas and
assumptions about data.
Analytic tractability:
statistics & tests
use equations in
closed form, each
tailored to the result
(e.g. mean).
Distributional
Assumptions:
(example) False
rejection of valid
data may occur for
fat tail distributions
when assumed
normal.
Statistical Data Algorithms:
Based on brute force re-sampling.
Confidence derives from sampling
theorems, algorithm properties,
and the statistic(s) applied to each
sample set.
(example)
Bootstrap
Algorithm:
Sampling with
replacement takes
statistic on data
subsets over many
iterations.
Assumptions re data
(usually) not required.
Sampling methods
may provide
distribution
insights.
Confidence
depends on the
statistic: still
difficult, but
flexible.
Field: Human-competitive prediction and decision
Barrier: Limited knowledge of how brains work
Insight: NN turns to previously-impractical EE/CS
Innovation: Ideas infeasible in 40’s-60’s are doable
“Biological” Neural Nets:
Parallel Distributed Processing
… gets a jumpstart from research
neglected by 60’s/70’s Symbolic
AI.
Practical implementations usually
based on CS, EE, OR methods.
Neural networks:
Good results from
early efforts drive
AI hype cycle,
inflated
expectations.
By 1991, backlog of
NN ideas tapped
out.
Problem:
We don’t know how
brains work, so
can’t make neural
nets “more brain-
like.”
CS, EE, OR are best
bets for new ideas.
Golden Oldies Triumphant:
Algorithms of the 50’s + CPUs of 90’s
Dynamic Programming, Monte Carlo
Methods, Convex Optimization, SVM
Popularized by their use in NN, their
variants are today’s core ML tools.
Feedback example:
Psychology
inspired the
Reinforcement
Learning(RL)
method, an ML
extensionofDynamic
Programming,nowa
frequentstapleofEE
& OR curricula.
The Connection:
Testing highly non-
linear methods
against problematic
data is an ideal
match for Statistical
Data Analysis
Algorithms.
*Statistical Data Analysis is the term coined by John Tukey, and used by the Efron & Tibshirani for continuity; however Statistical Data Analysis Algorithms is more accurate.

32.
The eve of disruption

33.
What is this thing called Computer Science?
Three Aspects:
• PROBLEM: Formally
states the input
elements and the goal
to achieve
• ALGORITHM: The
step sequence
required to achieve
the goal
• MECHANISM: The
system that
physically executes
the algorithm steps
Together, the three
aspects determine
hardness of a problem
+ solution pair in terms
of
• Time (number of
steps) and
• Space (memory)…
required to achieve
the goal.
Many problems are
intractable: they
require so many
steps to achieve a
perfect solution that
there isn’t enough
time and space in
the universe to find
it. Examples are
common:
• Making the perfect
chess move.
• Deciding what to wear
to the party next
week.
Intractable problems
can still be usefully
addressed using an
algorithm that finds
the best
approximate
solution for a given
amount of time and
space.
Computer Science is the study of mechanistic problem solving, starting
from three deeply intertwined aspects

34.
Machine Learning as Competitive
Advantage
• Google’s key competitive advantages in October 2005:
• The patent on Google’s ML-based PageRank algorithm was a
major barrier to entry for search engines thereafter
• Gmail featured an ML algorithm that eliminated spam, improved
user experience, and rapidly eroded competitors’ market share
• Google’s AdWords ML algorithm brought high revenue by rapidly
taking leadership of the online advertising market
• Google’s competitors were focused on “acquiring content.”
• The big technology push in corporate America? IT
Outsourcing.
• Behavioral biases 101: Computer Science had become framed by
most industries as a cost center, not as a source of innovation.

35.
Can the Google Disruption happen in finance?
There is at least one clue …
• The most popular desktop text on High Frequency Trading is
Irene Aldridge’s book of the same name.
• The finance material within is standard and thorough:
• risk adjusted return,
• alpha strategies / capacity,
• performance measurement,
• order books,
• detailed treatment of t-cost specific to microstructure
• What differentiates the book is its discussion of computer
architecture, networking, algorithms, and test procedures
• This is good solid CS 101 – no magical spells required
• It is a no-nonsense cookbook for industry disruption.
• ML in finance has no equivalent. Yet.

36.
CONCLUSIONS

37.
Anthropophoria: Heider-Simmel and AI
• Anthropomorphism is the great
mother of all behavioral biases
• Heider-Simmel led to the study of fundamental
attribution error (FAE), which motivated the work
of Kahneman and Tversky
• Artificial Intelligence seeks to make
computers do human-like things:
• Playing games
• Driving cars
• Recognizing cats
• … eliciting excitement and fear.
Whether asked pointed questions about the figure’s
behavior “as persons” or simply “what happened in the
movie,” subjects described the shapes as people with
vivid feelings and motivations.
When the film was shown in reverse from end-to-
start, the interactions of the figures made no obvious
sense, but separate groups of subjects again created
human behavioral interpretations of movement, and
described them equally vividly as did the subjects who
viewed the forward-running film.
Heider-Simmel (1943)
Subjects were shown a 3-minute film of moving
shapes interacting: collisions, pursuit, and avoidance
depicted highly intentional conflict among the trio.

38.
The Fundamental Law of AI
• Hype cannot diminish the significance of "deep" learning systems
outperforming humans on complex tasks.
• They are the newest in a long line of important AI developments over the past sixty
years, most of which are now forgotten.
• The forgotten AI Innovations did not disappear, or fail:
1. They became commonplace, …
2. the excitement wore off, and …
3. the world forgot that they were AI.
Autopilot, anti-lock brakes, optical character recognition, relational databases, and
chess-playing programs were once AI.
The Fundamental Law of Artificial Intelligence:
• AI is not viewed as AI after it becomes common, because
anthropophoria wears off.
• Corollary: The Fundamental Law of AI implies that AI failures
are remembered as AI.

39.
End
See Appendices

40.
Appendix 1: Endnotes
The title of this talk is taken from Probably Approximately Correct, the recent book by Leslie Valiant, which
greatly expands upon PAC Learning, a foundational theory of machine learning that first appeared in
Valiant, Leslie G. 1984. “A Theory of the Learnable.” Communications of the ACM 27 (11): 1134–1142.
This talk makes many mentions of computer-playing poker programs, particularly HUNL championship
program DeepStack. The Q-Group Fall Special Track on ML included dinner speaker Michael Bowling, who
heads the DeepStack team.
The https://deepstack.ai web site is accessible to both specialist and non-specialist viewers. It describes the
DeepStack project headed by Michael Bowling of U. Alberta, including gameplay footage, video lecture, and
pre-print of the cover article from the May 5, 2017 issue of Science. DeepStack is a program (one of two, the
other being CMU’s Libratus) that has significantly outperformed human champions at Heads Up No-Limit
Poker, a game of incomplete information, a property which places it in a category of difficulty far beyond Chess
or Go.
See also: Moravčík, et al, and Michael Bowling. 2017. “DeepStack: Expert-Level Artificial Intelligence in Heads-
up No-Limit Poker.” Science 356 (6337): 508. May 4, 2017.

41.
Appendix 2: Next Steps For Learning
Learning Machine Learning
First Step: Computer Science Overview
Christian and Griffiths, Algorithms to Live by.
This is a popular science and general readership book from which even professionals in the field can gain new insights. It draws
from life’s problems large and small: Where should I park? How long should I date before I marry? What’s the best way to sort my
laundry? From these life examples Computer Science is revealed as the study of problems, their characteristics, and the
algorithms that address them. The book illustrates hallmarks of what makes problems easy or hard, how solutions and algorithms
can employ tradeoffs to gain different types of efficiency, and how these ideas inform Economics, Political Decisions, and
Behavioral Psychology. Free of equations, suitable for audiobook, and recommended for all attendees.
Machine Learning for Quants, Hands On: Starting Point
James et al., An Introduction to Statistical Learning with Applications in R.
For quants intending to investigate Machine Learning, this is the preferred starting point because it approaches the field from a statistical
perspective that highlights similarities and differences with the foundations of Quant Finance practice. It is highly recommended to learn the
material via the online Stanford class based on this book (it is a live course, but recent semesters are archived for on-demand use). Course
and book are free, as are all software and data. NOTE: this class is taught by Tibshirani & Hastie through Stanford’s Lagunitas program, and
should not be confused with Andrew Ng’s Coursera Machine Learning class. Taking both isn’t a waste, but if forced to choose, Statistical
Learning is more compatible with Quant Finance.
(CONTINUED)

42.
Appendix 2: Next Steps For Learning
(cont)
Learning Machine Learning
The Full Reference: Statistical Learning
Hastie, Tibshirani, and Friedman, The Elements of Statistical Learning, 2ed.
The first edition of this book was the culmination of a ten-year transformation beginning in the early 1990’s that resulted in
both a rigorous grounding of ML in statistical method as well as firm establishment of Statistical Data Analysis first proposed
by John Tukey, where proofs about algorithms replace proofs about equations, reducing reliance on closed-form formulas
and restrictive assumptions on data distributions. While the authors were key figures in that movement, they carefully
cover the breadth of the statistical learning field and its contributors. Note: Intro to Statistical Learning is culled from this
book.
Supplement: Hastie, Tibshirani, & Wainwright, Statistical Learning with Sparsity.
Addresses the special case (or, perhaps, the common case) of too many candidate explanatory factors and not enough data.
Machine Learning for Quants, Hands On: 800 pages of Next Steps
Goodfellow, Bengio, and Courville, Deep Learning.
This book assumes minimum knowledge equivalent to understanding all concepts in Christian & Griffiths, Intro to Statistical
Learning, and undergrad-level calculus. Requires Python (learnable on-the-fly). Includes 200-page intro to linear algebra
and differential tensor calculus. Teaches math, statistical, and learning concepts, and requires large projects using
TensorFlow.

43.
Appendix 2B: Computer Age Reference
Statistics in the Computer Age
The New Classical Statistics
Efron, Brad and T. Hastie, Computer Age Statistical Inference.
Twenty-five years after Efron & Tibshirani’s seminal article, this book offers a full treatment of classical and modern
statistical inference. Rather than being focused primarily on Statistical Learning, it is intended to serve a modern age text on
statistics in general for teaching the basic curriculum.

44.
contact:
walter.tackett@nE12.com