International Journal on Natural Language Computing (IJNLC) Vol.11, No.

6, December 2022

UNSCRAMBLING CODES:
FROM HIEROGLYPHS TO MARKET NEWS

Emilio Barone1 and Gaia Barone2

1
Department of Economics and Finance, Luiss Guido Carli, Rome, Italy
2
School of Business, National College of Ireland, Dublin, Ireland

ABSTRACT
This paper reviews some of the steps that paved the way for the development of sentiment analysis (or
opinion mining), a technique apparently used by Jim Simons’ Medallion fund for scoring an ‘impossible’
performance: a 66% annual average rate of return in the 31 years between 1988 and 2018. Sentiment
analysis is a powerful tool that uses natural language processing (NLP), or computational linguistics, to
determine whether a text about a company is positive, negative or neutral and, in a final analysis, to dis-
cover stock price patterns. Humans have always used symbols to communicate, plainly or secretively.
Here we review some of the methods used in the past centuries, including Egyptians’ hieroglyphs, Julius
Caesar’s cipher, Fibonacci’s abbreviations, Leonardo da Vinci’s Mirror Writing, Mary Stuart’s code.
The intention is to describe some passages of the long journey made by human beings to arrive at the
current sophisticated IT tools for sentiment analysis.

KEYWORDS
Natural Language Processing, Hedge Funds, Cryptography

1. INTRODUCTION
What do Jean François Champollion, French archaeologist and Egyptologist, and James (Jim)
Harris Simons, mathematician and hedge fund manager, have in common (Figure 1)? They have
both been code breakers. Champollion was able to unscramble the message encrypted in the Ro-
setta stone. He had studied the Coptic language, which derived from ancient Egyptian, and this
knowledge made it possible for him to understand the Demotic section of the Rosetta stone. Jim
Simons received a bachelor’s degree in mathematics from MIT, a PhD in mathematics from UC
Berkeley and held a research position at Harvard. He later became a breaker of Russian codes at
the Institute for Defense Analysis (IDA), during the Cold War.

John Hull [1] cites Jim Simons in the chapter on Natural Language Processing (NLP) of his
book on Machine Learning in Business:
New data sources are becoming available all the time. One approach is to try and
be one step ahead of most others in exploiting these new data sources. Another is
to develop better models than those being used by others and then be very secretive
about it. Renaissance Technologies, a hedge fund, provides an example of the se-
cond approach. It has been amazingly successful at using sophisticated models to
understand stock price patterns. Other hedge funds have been unable to replicate
its success. The average return of its flagship Medallion fund between 1988 and
2018 was 66% per year before fees. This included a return of close to 100% in
2008 when the S&P 500 lost 38.5%. Two senior executives, Robert Mercer and Pe-
ter Brown, are NLP experts and have been running the company following the re-
tirement of the founder, Jim Simons, in 2009. (p. 197)
DOI: 10.5121/ijnlc.2022.11603 39
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

Source: Wikipedia
Figure 1 Jean-François Champollion and James (Jim) Harris Simons.

2. MEDALLION FUND
Gregory Zuckerman [2] in his bestselling book on The man who solved the market reported that
Simons quit Harvard in 1964 to join the Institute for Defense Analysis (IDA) “an elite research
organization that hired mathematicians from top universities to assist the National Security
Agency − the United States’ largest and most secretive intelligence agency − in detecting and
attacking Russian codes and ciphers. ... The IDA taught Simons how to develop mathematical
models to discern and interpret patterns in seemingly meaningless data.” (pp. 23-4).

While at IDA, Simons co-authored a paper on “Probabilistic Models for (and Prediction
of) Stock Market Behavior”. Detecting stock price patterns became one of his aims. However,
he apparently did not care about their interpretation:
“I don’t know why planets orbit the sun,” Simons told a colleague, suggesting one
needn’t spend too much time figuring out why the market’s patterns existed. “That
doesn’t mean I can’t predict them.” {[2], p. 151}
Whatever techniques used by Jim Simons, his results are astonishing (Appendix A):
There were compelling reasons I was determined to tell Simon’s story. A former
math professor, Simons is arguably the most successful trader in the history of
modern finance. Since 1988, Renaissance’s flagship Medallion hedge fund has
generated average annual returns of 66 percent, racking up trading profits of more
than $100 billion (see Appendix 1 for how I arrive at these numbers). No one in the
investment world comes close. Warren Buffett, George Soros, Peter Lynch, Steve
Cohen, and Ray Dalio all fall short (see Appendix 2). {[2], p. xvi}
Gross log-returns consistent with the data of Appendix A are reported in Table 1, together with
the correspondent data for the S&P 500 Total Return Index. The Medallion Fund (MF) outper-
formed the S&P 500 in 28 out of 31 years and never reported a loss. A $1 amount invested in
the Medallion Fund at the end of 1987 grew to $3,995,683 after 31 years!

Figure 2 reports the performance distributions of both investments (MF and S&P 500).
The average log-return, μ, of the Medallion Fund (49.0%) is 5 times greater than the average
log-return of the S&P 500 (9.7%), while the standard deviations, σ, are similar (18.7% and
17.0%, respectively). It will be impossible to beat this incredible performance!

40
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

Table 1 Medallion Fund vs. S&P 500 Total Return Index: gross log-returns per year.

End Medallion S&P 500 End Medallion S&P 500

of year Fund Total Return of year Fund Total Return
1988 15.1% 15.4% 2005 45.6% 4.8%
1989 1.0% 27.5% 2006 61.0% 14.7%
1990 57.5% -3.2% 2007 86.1% 5.3%
1991 43.4% 26.6% 2008 92.5% -46.2%
1992 38.5% 7.3% 2009 55.7% 23.5%
1993 43.1% 9.6% 2010 45.4% 14.0%
1994 66.0% 1.3% 2011 53.7% 2.1%
1995 42.5% 31.9% 2012 45.0% 14.8%
1996 36.7% 20.7% 2013 63.6% 28.1%
1997 27.4% 28.8% 2014 56.0% 12.8%
1998 45.2% 25.1% 2015 52.7% 1.4%
1999 30.5% 19.1% 2016 52.2% 11.3%
2000 82.5% -9.5% 2017 61.7% 19.7%
2001 44.9% -12.7% 2018 56.8% -4.5%
2002 41.3% -25.0%
2003 36.5% 25.2% μ 49.0% 9.7%
2004 40.2% 10.3% σ 18.7% 17.0%
Note: The linear regression of Medallion Fund on S&P 500, based on normalized log-returns, gives α = 0 (t-
stat = 0) and β = −0.5 (t-stat = −3.3).

12 0.40
S&P 500 Total Return
Medallion Fund 0.35
10
N(9.7%, 17.0%)
0.30
Number of observations

N(49.0%, 18.7%)
8
0.25
Normal density

6 0.20

0.15
4
0.10
2
0.05

0 0.00

Source: Zuckerman {[[2], Appendix 1} and Yahoo Finance (^SP500TR).

Figure 2 Medallion Fund vs. S&P 500 Total Return.

2.1 The Impossible

“The Impossible” is the title of a movie based on the 2004 Indian Ocean tsunami, an extreme
observation among the low-frequency high-severity natural disasters. Sometimes very rare phe-
nomena as the S&P 500 fall of October 19, 1987 (-20.5%, from 282.70 to 224.84) and March

41
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

16, 2020 (-12.0%, from 2,711.02 to 2,386.13) are called “6-sigma” or “six standard deviation”
events. Actually, if X is a standardized normal variable, the probability of X > 6.4
[=NORM.S.INV(1-1E-10) in Excel™] is equal to 0.0000000001, i.e. 1 out of 10 billion.

What is the probability of earning a 49.0% average annual log-return for 31 consecutive
years without never reporting a loss?

2.2 Leverage and Roulette Wheel

Stellar performances are not possible without leverage. A brilliant example of leverage’s effect
on performance has been offered by Mark Rubinstein [3]:
Let’s suppose, for the sake of this discussion, that the roulette wheel only has num-
bers one through 36 on it, and that it doesn’t have a zero or a double zero. If you
place one chip on each of the numbers from one through 35, you will win a small
amount most of the time. Now suppose you borrowed most of the money to finance
the 35 chips you bet. You’d find that most of the time you’d earn a high rate of re-
turn on your capital, but that, on average, one out of 36 times you would lose it all.
... The general problem with one–35 and 36 is one of evaluating investment strate-
gies. A lesson we can draw is that when we look at the performance of an invest-
ment manager, we now have to ask ourselves, Is this a strategy that wins 35 out of
36 times, but really blows it the 36th time? Because of this problem, it’s easy to be
deceived by a manager’s track record. (pp. 3-4)

2.3 Natural Language Processing

Our a priori is that there is much more than “good luck” behind the stellar performance of Me-
dallion Fund: best human capital, massive usage of computer power, and natural language pro-
cessing (NLP). Robert Mercer and Peter Brown, who joined forces with Jim Simons in 1993,
were both working at IBM “searching for ways to get computers to do a better job transcribing
speech into text and even translate language” {Zuckerman [2], p. 4}. The search for a common
language, or binarisation, was the ultimate goal. Probably, natural language processing (NLP) or
computational linguistics has been the key breakthrough that helps to explain most of MF’s per-
formance. Much of the big data generated in the world is in the form of written or spoken
words. Quotes and volumes are basic statistics that summarize a great deal of information, but
market news is just as important. The rules of language are difficult to communicate to a ma-
chine and words can have several meanings. To crack the market’s code was not an easy job.

In the text that follows we will describe some passages of the long journey made by
human beings to communicate with each other, in a manifest or secret way. We will deal with
Egyptian Hieroglyphs (§3), Julius Caesar’s cipher (§4), Fibonacci’s Liber Abaci (§5), Leonardo
da Vinci’s Mirror Writing (§6), Mary Stuart’s Code (§6), Monte Carlo Descrambling (§8),
Shazam and Google Lens (§9), Kaggle (§10). Then we will conclude (§11).

3. EGYPTIAN HIEROGLYPHS
3.1 Rosetta Stone
The Rosetta Stone is a black granite stele inscribed with three versions of a decree issued in
Memphis, Egypt, in 196 BC during the Ptolemaic dynasty on behalf of King Ptolemy V
Epiphanes. It was discovered in July 1799 by French officer Pierre-François Bouchard during
the Napoleonic campaign in Egypt, while digging the foundations of an addition to a fort in the
Nile Delta, near el-Rashid (Italianized in Rosetta, which means “little rose”). It is an irregularly
shaped grey and pink granite stone [3 feet 9 inches (114 cm) long and 2 feet 4.5 inches (72 cm)
wide], exhibited in the British Museum since 1802 (Figure 3).

42
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

Source: © The Trustees of the British Museum.

Figure 3 Rosetta Stone.

The Rosetta Stone bears a priestly decree concerning Ptolemy V in three blocks of text:
Hieroglyphic (the “language of the gods”, 14 lines), Demotic (the “language of acts”, 32 lines)
and Ancient Greek (53 lines). Translation of the Ancient Greek was relatively easy. Translation
of Demotic (the ancient Egyptian script preceding Coptic) was harder, but the proper names
(Ptolemy, Alexander and Alexandria) were quickly deciphered. Translation of hieroglyphic text
was even harder, but it was soon established that names of kings or pharaohs were contained
within elongated ovals (“cartouches”).

43
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

(a) original from right to left

(b) transcribed from left to right

Source: British Museum (original).

Figure 4 The name Ptolemy in the Rosetta Stone.

There were six cartouches on the Rosetta Stone. According to the Greek translation,
these cartouches clearly had to contain the name Ptolemy (Ptolemaios, in Greek). Three of them
looked as in Figure 4 and featured the name Ptolemy along with an Egyptian honorific: “Ptole-
my, living for ever, beloved of Ptah” {Robinson [4], p. 125}. The other three looked like this:

Hieroglyph
𓊪𓊪 𓏏𓏏 𓍯𓍯 𓃭𓃭 𓐝𓐝 𓇌𓇌 𓋴𓋴
Champollion’s reading P T O L M E S

The hieroglyphs of ancient Egypt are one of the earliest forms of writing. They can be read from
right to left or left to right. You can distinguish the direction in which the text is to be read be-
cause the human or animal figures always face towards the beginning of the line. Also the upper
symbols are read before the lower. There are no spaces between words, line breaks or punctua-
tion.

3.2 Google Translator

On July 15, 2020 (exactly 221 years after the Rosetta Stone was discovered) Google has
launched a hieroglyphics translator that uses machine learning to decode ancient Egyptian lan-
guage: https://artsexperiments.withgoogle.com/fabricius/en (Table 2).

Google created tools and models for the three phases of the algorithm:
1) Extraction - taking hieroglyphic script and sequences from source images and creating
workable facsimiles;
2) Classification - training a neural network to correctly identify over 1000 hieroglyphs;
3) Translation - matching sequences and blocks of text to available dictionaries and published
translations.

44
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

Table 2 Ancient Egypt Hieroglyphs.

AEO A B CH D F PH V

𓄿𓄿 𓂝𓂝 𓃀𓃀 𓍿𓍿 𓂧𓂧 𓆑𓆑
vulture 1313F forearm 1309D foot 130C0 hobble rope 1337F hand 130A7 horned viper 13191

G H H IY J (G) K (C)

𓎼𓎼 𓎛𓎛 𓉔𓉔 𓇋𓇋 𓆓𓆓 𓎡𓎡
pot stand 133BC rope 1339B shelter 13254 reed leaf 131CB cobra 13193 basket 133A1

K (C) LR M N OUW P

𓏘𓏘 𓂋𓂋 𓅓𓅓 𓈖𓈖 𓅱𓅱 𓊪𓊪
hillside 133D8 open mouth 1308B owl 13153 water 13216 quail chick 13171 stool 132AA

S (C) SH T TH TH Z

𓋴𓋴 𓈙𓈙 𓏏𓏏 𓐍𓐍 𓄡𓄡 𓊃𓊃
folded cloth 132F4 lake 13219 bread loaf 133CF (unknown) 1340D cow’s belly 13121 door bolt 13283

Note: 1,079 hieroglyphs are available at https://fonts.google.com/noto/specimen/Noto+Sans+Egyptian+Hieroglyphs.

4. JULIUS CAESAR’S CIPHER

Julius Caesar (12 July 100 BC - 15 March 44 BC) was well ahead of his time. At a time when
illiteracy was the norm, he preferred to communicate with his generals by means of a secret
code, Caesar’s cipher. Suetonius, the private secretary to the Emperor Hadrian, tells us that Jul-
ius Caesar used a mono-alphabetical cipher where each letter was replaced by the letter that fol-
lowed it by three places in the alphabet, for his confidential correspondence:

Extant et ad Ciceronem item ad There are also letters of his to Cic-

familiares domesticis de rebus in ero, as well as to his intimates on
quibus si qua occultius perferenda private affairs, and in the latter, if
erant per notas scripsit id est sic he had anything confidential to say,
structo litterarum ordine ut nul- he wrote it in cipher, that is, by so
lum verbum effici posset: quae si changing the order of the letters of
quis investigare et persequi volet the alphabet, that not a word could
quartam elementorum litteram id be made out. If anyone wishes to
est D pro A et perinde reliquas decipher these, and get at their
commutet. meaning, he must substitute the
fourth letter of the alphabet, name-
ly D, for A, and so with the others.

Source: Suetonius [5], De vita Caesarum, Ch. 1, §56.

45
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

A similar cipher was used by the Emperor Augustus (23 September 63 BC - 19 August AD 14):

Orthographiam, id est formulam He does not strictly comply with or-

rationemque scribendi a gram- thography, that is the say the theo-
maticis institutam, non adeo cus- retical rules of spelling laid down
todit ac videtur eorum potius se- by the grammarians, seeming to be
qui opinionem, qui perinde rather of the mind of those who be-
scribendum ac loquamur exis- lieve that we should spell exactly as
timent. Nam quod saepe non lit- we pronounce. Of course his fre-
teras modo sed syllabas aut per- quent transposition or omission of
mutat aut praeterit, communis syllables as well as of letters are
hominum error est. slips common to all mankind.
Nec ego id notarem, nisi mihi I should not have noted this, did it
mirum videtur tradidisse aliquos, not seem to me surprising that some
legato eum consulari successo- have written that he cashiered a
rem dedisse ut rudi et indocto, consular governor, as an unculti-
cuius manu "ixi" pro ipsi scrip- vated and ignorant fellow, because
tum animadverterit. he observed that he had written ixi
for ipsi.
Quotiens autem per notas scribit, Whenever he wrote in cipher, he
B pro A, C pro B ac deinceps wrote B for A, C for B, and the rest
eadem ratione sequentis litteras of the letters on the same principle,
ponit; pro X autem duplex A. using AA for X.

Source: Suetonius [5], De vita Caesarum, Ch. 2, §88.

Figure 5 shows the Alberti cipher disk, an enciphering and deciphering tool developed in 1470
by Leon Battista Alberti. The device consists of two concentric circular plates mounted one on
top of the other.

Source: Wikipedia.
Figure 5 Alberti cipher disk.

46
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

5. FIBONACCI’S LIBER ABACI

Before Gutenberg invented the printing press, around 1440, texts were handwritten. Even if
there is no secret code to decrypt, it is sometimes hard to “translate” characters in machine-
readable text. This is the case of Liber Abaci, the book written in 1202 by Leonardo Pisano,
known today to mathematicians and scientists all over the world by the name Fibonacci (Table 3).

Table 3 Fibonacci’s Liber Abaci.

...

Manuscript| Liber Abaci, Leonardo Pisano, 1202 | Codice Magliabechiano, C. I, 2616, Badia Fiorentina, n. 73.

Incipit liber Abaci Compositus a leonardo filio Bonacij Pisano In Anno. M° cc° ij.°

S CRIPSISTis mihi domine mi magister Michael Scotte, summe philosophe, vt librum de numero, quem dudum
composui, uobis transcriberem: vnde uestrae obsecundans postulationi, ipsum subtiliori perscrutans Indagine ad
uestrum honorem et aliorum multorum utilitatem correxi. In cuius correctione quedam necessaria addidj, et quedam
superflua resecaui. In quo plenam numerorum doctrinam edidj, iuxta modum indorum, quem modum in ipsa scientia
prestantiorem elegi.
...
Incipit primum capitulum. Nouem figure indorum he sunt
9 8 7 6 5 4 3 2 1.
Cvm his itaque nouem figuris, et cum hoc signo 0, quod arabice zephirum appellatur, scribitur quilibet numerus, ut
inferius demonstratur.
“Decrypted” Latin | Boncompagni, B. [6], Il Liber Abbaci di Leonardo Pisano | Rome, 1857.

Here begins the Book of Calculation Composed by Leonardo Pisano, Family Bonaci, In the Year 1202.
You, my Master Michael Scott, most great philosopher, wrote to my Lord about the book on numbers which some
time ago I composed and transcribed to you; whence complying with your criticism, your more subtle examining cir-
cumspection, to the honor of you and many others I with advantage corrected this work. In this rectification I added
certain necessities, and I deleted certain superfluities. In it I presented a full instruction on numbers close to the
method of the Indians, whose outstanding method I chose for this science.
...
Here Begins the First Chapter. The nine Indian figures are:
9 8 7 6 5 4 3 2 1.
With these nine figures, and with the sign 0 which the Arabs call zephir any number whatsoever is written, as is
demonstrated below.
English | Sigler, L. E. [7], Fibonacci’s Liber Abaci, pp. 15 and 17, 2002.

47
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

6. LEONARDO DA VINCI’S MIRROR WRITING

Leonardo da Vinci (15 April 1452 – 2 May 1519) wrote backward, from right to left, most of
his personal notes, in an attempt to keep them illegible to anyone other than him. For instance,
he wrote:

[Leonardo, in Italian, encrypted] [Leonardo, in Italian, decrypted]

ehc olopecsid leuq ’e otsirt   tristo e’ quel discepolo che
_ ortseam ous li aznava non non avanza il suo maestro _

flipped to

[Text in English, encrypted] [Text in English, decrypted]

taht lipup eht si roop   poor is the pupil that
_ retsam sih ssaprus ton seod does not surpass his master _
Source: 1493, Forster III, South Kensington Museum III 24b, http://t.co/ON3xNdrOhb.

7. MARY STUART CODE

On February 8, 1587, Mary Queen of Scots was beheaded for high treason after nineteen years
of imprisonment in England. She was found guilty of her involvement in the Babington plot, a
conspiracy to assassinate Elizabeth I organized by Anthony Babington. Mary's encoded corre-
spondence had been decrypted {Singh, [8]}. The alphabet used is shown in Table 4.

Table 4 Mary Stuart Code | Alphabet

A B C D E F G H I J K

L M N O P Q R S T U V

W X Y Z and for but with that if as

of the by so not when from this is in say

me my I you what where which there send receive pray

Note: The alphabet contains 5 blank characters that can be used as space
or word separator (this makes the frequency analysis more complex):

Source: https://www.dcode.fr/mary-stuart-code.

48
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

8. MONTE CARLO DESCRAMBLING

The first step to decrypt a code is to guess the text’s language. Then a frequency analysis can be
applied. The current version of Google’s translator, launched in April 2006, can recognize 108
languages, from Afrikaans to Xhosa, Yiddish, Yoruba, and Zulu {WU, Y., et al. [10]}.

The statistical distribution of letters in the English language, derived from some classic
books, is reported in Table 5.

Table 5 Statistical Distributions of English Text: Letter Frequencies.

a 6.51738% h 4.92888% o 5.96302% v 0.82903%

b 1.24248% i 5.58094% p 1.37645% w 1.71272%
c 2.17339% j 0.09033% q 0.08606% x 0.13692%
d 3.49835% k 0.50529% r 4.97563% y 1.45984%
e 10.41442% l 3.31490% s 5.15760% z 0.07836%
f 1.97881% m 2.02124% t 7.29357% Space 19.18182%
g 1.58610% n 5.64513% u 2.25134% 100.00000%

Source: Art Owen [11], Stat 362 (Monte Carlo Methods), Stanford University, Fall 2008.
Note: The above statistics are derived from these classics (available at www.literature.org): Origin of
Species (Charles Darwin), The Voyage of the Beagle (Charles Darwin), Jane Eyre (Charlotte Bronte),
Wuthering Heights (Emily Bronte), Tarzan of the Apes (Edgar Rice Burroughs), The Return of Tarzan
(Edgar Rice Burroughs), Paradise Lost (John Milton). All upper-case letters were converted to lower-
case. All numbers and all special characters were removed. A carriage return is treated as a space. There
are exactly 5,086,936 characters in these files.

The Monte Carlo approach also requires the probabilities of adjacent letters, i.e. the (first-order)
letter transition probabilities:

Prob ≡ Prob(x +1 = j | x = i) for 1 ≤ i, j ≤ 27.

i→j

A real-life example of Monte Carlo descrambling is shown by Persi Diaconis {[9], pp. 1-3}:
One day, a psychologist from the state prison system showed up with a collection
of coded messages. ... The problem was to decode these messages. Marc [Coram]
guessed that the code was a simple substitution cipher, each symbol standing for a
letter, number, punctuation mark or space. ... I like this example because a) it is
real, b) there is no question the [Monte Carlo] algorithm found the correct answer,
and c) the procedure works despite the implausible underlying assumptions. In
fact, the message is in a mix of English, Spanish and prison jargon. The plausibility
measure is based on first-order transitions only. A preliminary attempt with single-
letter frequencies failed.

9. SHAZAM AND GOOGLE LENS

Binarisation is the preprocessing technique used to transform characters, sounds and images in
binary digits (0, 1). A great part of all the available information is binarised. Optical character
recognition (OCR) has made the electronic conversion of typed, handwritten or printed text into
machine-encoded text possible. Optical music recognition (OMR) has the same aim, but is
based on printed sheet music. Shazam, owned by Apple, uses the microphone on the device to
identify music, movies, advertising, and television shows, while Google Lens identifies images
in a few instants (Table 6).

49
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

Table 6 Music and Image Recognition: Shazam and Google Lens.

Marilyn Monroe Romy Schneider Romy Schneider Michel Piccoli Humphrey Bogart Jane Birkin
Arthur Miller Peter O'Toole Alain Delon Romy Schneider Lauren Bacall Serge Gainsbourg

Alain Delon Steve McQueen Frank Sinatra Grace Kelly

Recognized Unrecognized
Nathalie Delon Neile Adams Ava Gardner Ray Milland

Gena Rowlands Maggie Gyllenhaal Ethan Hawke

Recognized Unrecognized Unrecognized
John Cassavetes Peter Saarsgard Winona Ryder

George Harrison Jane Asher Mel Ferrer Jacqueline Kennedy Jim Morrison
Unrecognized
Pattie Boyd Paul McCartney Audrey Hepburn John Fitzgerald Kennedy Pamela Courson

Errol Flynn Isabella Rossellini Jacques Charrier François Truffaut

Recognized Unrecognized
Olivia de Havilland David Lynch Brigitte Bardot Catherine Deneuve

Gregory Peck Marcello Mastroianni Richard Burton Audrey Hepburn Audrey Hepburn Gad Elmaleh
Ingrid Bergman Anna Karina Elizabeth Taylor Mel Ferrer Mel Ferrer Audrey Tautou

Neile Adams Natalie Wood Jack Nicholson Anna Karina Romy Schneider Monica Vitti
Steve McQueen Robert Redford Anjelica Huston Jean-Luc Godard Alain Delon Michelangelo Antonioni

Jean-Paul Belmondo James Dean

Unrecognized Unrecognized Unrecognized Recognized
Jean Seberg Anna Maria Pierangeli

Marilyn Monroe Alain Delon Alain Delon Claude Mann Cary Grant
Recognized
Eli Wallach Romy Schneider Nathalie Delon Jeanne Moreau Katharine Hepburn

Monica Bellucci Chiara Mastroianni Märta Torén

Vincent Cassel Benicio del Toro Dana Andrews

Source: YouTube (https://youtu.be/ck57LnYScNQ), Shazam (Burt Bacharach, Live in Sidney, 2008) and
Google Lens.

10. KAGGLE
A huge amount of market news available on social media and elsewhere needs to be worked out
to discover if they will have a positive, negative, or neutral effect on a particular company. A
good starting point to learn Natural Language Processing (NLP) is to participate in a competi-
tion organized by Kaggle, a subsidiary of Google. One of the Kaggle’s “Getting Started” com-
petitions is titled “Natural Language Processing with Disaster Tweets: Predict which Tweets are
about real disasters and which ones are not.” All of the work can be done in Kaggle’s free, no-
setup, Jupyter Notebooks environment, where you can run Python code.

50
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

The competition is ongoing. Table 7 shows a few steps of the current leader’s code.

Table 7 Kaggle’s Competition: Natural Language Processing with Disaster Tweets.

A few steps of the code used by the current leader (Shahules).

Importing required Libraries.
Using TensorFlow backend. [TensorFlow: an open-source machine learning platform developed by Google]
Loading the data and getting basic idea [Training set: 7613 rows and 5 columns; Test set: 3263 rows and 4 columns]
Class distribution [There are more tweets with class 0 (no disaster) than class 1 (disaster tweets)]
Exploratory Data Analysis of tweets [at character level, word level and sentence level]
Number of characters in tweets, number of words in a tweet, average word length in a tweet, common stopwords
in tweets [stopwords: words that are so commonly used that they carry very little useful information (the, in, of, a,
to, and, on, for, is, at, ...)], analyzing punctuations [=> - : ? . | + # ! ] ) ’ ; / = @ ~ * ...], Ngram analysis [Ngram: a
sequence of N words]
Data Cleaning [Removing URLs, HTML tags, Emojis, Punctuations. Spelling Correction]
GloVe for Vectorization [GloVe: an unsupervised learning algorithm for obtaining vector representations for words.]
Number of unique words: 20342
Baseline Model
Source: www.kaggle.com/c/nlp-getting-started/overview, www.kaggle.com/shahules/basic-eda-cleaning-and-glove.

51
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

11. FINAL REMARKS

The average annual log return of Medallion Fund, the hedge fund of Jim Simons’ Renaissance
Technologies, has been 49.0% in the 31 years from 1988 to 2018. This stellar performance is in
strong contrast to the Efficient Market Hypothesis (EMH). If markets were informationally effi-
cient, none could perform as Medallion Fund (MF). So, what was the MF’s key to success? It is
not enough to process numerical data, such as stock prices and volumes. The MF’s key to suc-
cess seems to be Natural Language Processing (NLP):
“There are patterns in the market,” Simons told a colleague, “I know we can find
them.” {[2], p. 5}.
The daunting goal of NLP is to successfully perform human like language processing. In order
to make investment/disinvestment decisions, people rely on news reports about companies, on
quarterly earnings calls to analysts, on opinions posted on the web, etc. Baker and Wurgler [12]
were among the first researchers to study how investor sentiment affects the cross-section of
stock returns. Zhang and Skiena [13] exploited blog and news sentiment to construct market
neutral portfolios that produced impressive returns. We believe that information retrieval tech-
niques, as those shown in John and Govilkar [14], will make the difference between failure or
success in the world of investing.

REFERENCES
[1] Hull, J. C., Machine Learning in Business: An Introduction to the World of Data Science, 3rd ed.,
KDP, May 26th, 2021.
[2] Zuckerman, G., The man who solved the market: How Jim Simons launched the quant revolution,
Penguin Random House, 2019.
[3] Rubinstein, M., “The World According to Mark Rubinstein: Interview”, Derivatives Strategy Maga-
zine, July 1999.
[4] Robinson, A., “Cracking the Egyptian Code: The Revolutionary Life of Jean-François Champol-
lion”, Thames & Hudson, 2018.
[5] Suetonius, G., The Twelve Caesars: The Lives of the Roman Emperors, J. C. Rolfe (Trans.). St. Pe-
tersburg, FL: Red and Black Publications, 2008.
[6] Boncompagni, B., Il Liber Abbaci di Leonardo Pisano, Rome, 1857.
[7] Leonardo da Pisa, Liber Abaci, 1202. See SIGLER, L. E., Fibonacci’s Liber Abaci - A Translation
into Modern English of Leonardo Pisano’s Book of Calculation, Springer, 2002.
[8] Singh, S., The Code Book: The Science of Secrecy from Ancient Egypt to Quantum Cryptography,
Fourth Estate, 1999.
[9] Diaconis, P., “The Markov Chain Monte Carlo Revolution,” Bulletin of the American Mathematical
Society, 46(2), 179-205, 2009.
[10] Wu, Y., et al., “Google’s neural machine translation system: Bridging the gap between human and
machine translation,” https://arxiv.org/pdf/1609.08144.pdf (2016).
[11] Owen, A., “Monte Carlo theory, methods and examples,” https://artowen.su.domains/mc/.

[12] Baker, M., and Wurgler, J., “Investor Sentiment and the Cross-Section of Stock Returns”, Journal of
Finance, Vol. 61, No. 4, 2006.

[13] Zhang, W., and Skiena, S., “Trading Strategies to Exploit Blog and News Sentiment”, Proceedings
of the Fourth International AAAI Conference on Weblogs and Social Media, 2010.

[14] John, R., and Govilkar, S., “Information Retrieval Technique for Web Using NLP”, International
Journal on Natural Language Computing, Vol. 6, No. 5, 2017.

52
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

AUTHORS
Emilio Barone is a Stephen A. Ross Professor of Financial Economics, Intesa Sanpaolo
Chair, in the Department of Economics and Finance, Luiss Guido Carli University of
Rome. He has been an Executive Director at Intesa Sanpaolo, Head of the Financial
Risks Analysis Department and Head of Research at Istituto Mobiliare Italiano (I.M.I.).
He has previously worked as an economist in the Research Department of the Bank of
Italy. He has followed postgraduate studies at Yale University after receiving his
B.S./M.S. in Economics from “La Sapienza” University of Rome. He is the author of
many articles on derivatives pricing and risk measurement / management.

Gaia Barone is an Assistant Professor in “Economics and Finance” at the School of

Business at National College of Ireland. She is Program Director for the M.Sc. in Fi-
nance. Before September 2018, she was an Assistant Professor in “Mathematical
Methods for Economics”, and Chair of “Financial Mathematics” and “Quantitative
Methods for Management” at Luiss Guido Carli University of Rome. In July 2011 she
received her Ph.D. in Money and Finance from “Tor Vergata” University of Rome and
in June 2009 her M.Sc. in Financial Mathematics from Stanford University. She has
published two books on arbitrages and articles on derivatives and credit risk.

53
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

APPENDIX A
THE MAN WHO SOLVED THE MARKET BY GREGORY ZUCKERMAN
Appendix 1
Net Management Performance Returns Size Medallion
Returns Fee* Fee Before Fees of Fund Trading Profits**
1988 9.0% 5% 20% 16.3% $20 million $3 million
1989 -4.0% 5% 20% 1.0% $20 million $0 million
1990 55.0% 5% 20% 77.8% $30 million $23 million
1991 39.4% 5% 20% 54.3% $42 million $23 million
1992 33.6% 5% 20% 47.0% $74 million $35 million
1993 39.1% 5% 20% 53.9% $122 million $66 million
1994 70.7% 5% 20% 93.4% $276 million $258 million
1995 38.3% 5% 20% 52.9% $462 million $244 million
1996 31.5% 5% 20% 44.4% $637 million $283 million
1997 21.2% 5% 20% 31.5% $829 million $261 million
1998 41.7% 5% 20% 57.1% $1.1 billion $628 million
1999 24.5% 5% 20% 35.6% $1.54 billion $549 million
2000 98.5% 5% 20% 128.1% $1.9 billion $2,434 million
2001 33.0% 5% 36% 56.6% $3.8 billion $2.149 million
2002 25.8% 5% 44% 51.1% $6.24 billion $2.676 billion
2003 21.9% 5% 44% 44.1% $5.09 billion $2.245 billion
2004 24.9% 5% 44% 49.5% $5.2 billion $2.572 billion
2005 29.5% 5% 44% 57.7% $5.2 billion $2.999 billion
2006 44.3% 5% 44% 84.1% $5.2 billion $4.374 billion
2007 73.7% 5% 44% 136.6% $6.2 billion $7.104 billion
2008 82.4% 5% 44% 152.1% $5.2 billion $7.911 billion
2009 39.0% 5% 44% 74.6% $5.2 billion $3.881 billion
2010 29.4% 5% 44% 57.5% $10 billion $5.750 billion
2011 37.0% 5% 44% 71.1% $10 billion $7.107 billion
2012 29.0% 5% 44% 56.8% $10 billion $5.679 billion
2013 46.9% 5% 44% 88.8% $10 billion $8.875 billion
2014 39.2% 5% 44% 75.0% $9.5 billion $7.125 billion
2015 36.0% 5% 44% 69.3% $9.5 billion $6.582 billion
2016 35.6% 5% 44% 68.6% $9.5 billion $6.514 billion
2017 45.0% 5% 44% 85.4% $10 billion $8.536 billion
2018 40.0% 5% 44% 76.4% $10 billion $7.643 billion
39.1% 66.1% $104.530.000.000
average average returns total
net returns before fees trading profits
* Fees are charged by the Medallion fund to its investors, which in most years represents the firm's own employees
and former employees.
** Gross returns and Medallion profits are estimates − the actual number could vary slightly depending on when the
annual asset fee is charged, among other things. Medallion's profits are before the fund's various expenses.
Average Annual Returns: 66.1% gross, 39.1% net
The above profits of $104.5 billion represent those of the Medallion fund. Renaissance also profits from three hedge
funds available to outside investors, which managed approximately $55 billion as of April 30, 2019. (Source: Medal-
lion annual reports; investors)

54
 International Journal on Natural Language Computing (IJNLC) Vol.11, No.6, December 2022

Appendix 2

Returns Comparison
Investor Key Fund/Vehicle Period Annualized Returns*
Jim Simons Medallion Fund 1988-2018 39.1%
George Soros Quantum Fund 1969-2000 32.0%†
Steven Cohen SAC 1992-2003 30.0%
Peter Lynch Magellan Fund 1977-1990 29.0%
Warren Buffett Berkshire Hathaway 1965-2018 20.5%‡
Ray Dalio Pure Alpha 1991-2018 12.0%
* All returns are after fees.
† Returns have fallen in recent years as Soros has stopped investing money for others.
‡ Buffett averaged 62% gains investing his personal money from 1961 to 1957, starting with
less than $10,000, and saw average gains of 24.3$% for a partnership managed from 1957 to
1969.

Appendix 2 does not report the performance of Princeton/Newport Partners, the hedge fund of
Edward Thorp. However, Zuckerman mentions it elsewhere {[2], pp. 127-9}:

During the 1970s, Thorp helped lead a hedge fund, Princeton/Newport Partners,
recording strong gains and attracting well-known investors ... by the late 1980s,
Thorp’s fund stood at nearly $300 million, dwarfing the $25 million Simons’s Me-
dallion fund was managing at the time ... Over its nineteen-year existence, the
hedge fund featured annual gains averaging more than 15 percent (after charging
investors various fees), topping the market’s returns over that span.

55