MONEY WITHOUT MASTERS AN EXPLORATION OF VOLATILITY IN CRYPTO MARKETS

MONEY WITHOUT MASTERS: AN EXPLORATION OF VOLATILITY IN

CRYPTO MARKETS

CHAPTER I

I. INTRODUCTION

“Stay away from it. It’s a mirage, basically. In terms of cryptocurrencies, generally, I can

say almost with certainty that they will come to a bad ending.” – Warren Buffett

“I do think Bitcoin is the first [encrypted money] that has the potential to do something

like change the world.” – Peter Thiel, co-founder of PayPal

1.1 Background

Are cryptocurrencies really currencies? Defined as “a form of digital currency that can

serve as a medium of exchange and a store of value … but lacking (sic) the backing of

any central authority or government (Cetina & Hoffman, 2022), cryptocurrencies are

disrupting the financial services industry (Cheung et al., 2015). Once used only by

technology enthusiasts and dark web actors, cryptocurrencies have garnered widespread

adoption with the potential to reduce, or even eliminate, the use of local fiat currencies

(Taskinsoy, 2019). In 2021, cryptocurrencies surpassed a global market capitalization of

$2 trillion representing approximately 4% of global gross domestic product (GDP)

(Chinchalkar, 2021). Yet, there is little consensus that digital currency is to be treated like

traditional currency. Markets, governments, even cryptocurrency users have struggled to

assign an existing label to them (Vigna & Casey, 2016). Some see cryptocurrencies as

securities, to be traded like corporate stocks (Hacker & Thomale, 2018). Others view

cryptocurrencies as a form of digital asset, a digital representation of their ownership of

“cyber space” (Bentov et al., 2019). Despite the lack of a discernable definition, millions

of people, organizations, and now countries have progressed toward cryptocurrency

adoption.

In 2008, when the Bitcoin whitepaper was published, cryptocurrencies exploded

as a market force with the potential for applications ranging from expedited check

processing to national currencies (for example, in El Salvador). The Libra Association,

with members including Facebook and Mastercard, considered the development of its

own cryptocurrency that would replace the need for fiat currencies – a project that never

actually launched (Kharif, 2022; The Libra Association, 2019). However, other consortia

have since emerged to develop cryptocurrencies which are still underway (Phillips,

2022). Still, great debate remains as to whether cryptocurrencies are money, securities,

commodities, or their own thing entirely. Furthermore, cryptocurrency price volatility

may be preventing cryptocurrencies from becoming a mainstream mechanism for

payments.

In 2017, it was estimated that there were between 2.9 and 5.8 million unique,

active Bitcoin users while there were only a handful of corporations accepting Bitcoin

payments (Hileman & Rauchs, 2017). By 2022, that number ballooned to more than 180

million users (Howarth, 2022). Cryptocurrencies represent a new era in transaction

dynamics. This study seeks to understand how various characteristics of

cryptocurrencies

– and the blockchain technology that supports them – impacts their volatility.

Cryptocurrency prices and price volatility are the subject of much speculation

amongst investors. Some liken cryptocurrencies to the “tulip fever” during the Dutch

Golden age when people paid exorbitant prices for tulip bulbs that had no value. Others

see cryptocurrencies as the next iteration of finance – one where centralized banks and

other entities no longer control the economy. Several studies have attempted to explain

the fluctuations in cryptocurrency prices with the results largely explaining what does not

impact prices. For example, Baur, Dimpfl, and Kuck (2018) conducted a study to

understand whether gold spot prices had any predictive relationship to Bitcoin price

volatility but found no such relationship. To date, relatively little is known about the

behavior of digital assets as their volatility does not appear in alignment with the

behavior of currencies, securities, or commodities. Stablecoins, a form of digital asset

that are designed to mimic the price performance to the specific fiat currencies to which

they are “pegged” have received virtually no attention in academic literature. Further, few

studies have looked at internal factors, such as blockchain performance variables, to see

if predictive relationships might exist between them and cryptocurrency prices. A few

studies have examined individual blockchain performance variables, such as hash rate,

for their predictive relationships to cryptocurrency prices, but none have looked at a

comprehensive set of these variables.

This study was conducted in three parts. Study 1a predicts the volatility of six

cryptocurrencies using existing asset classes (i.e., securities, commodities, and fiat

currencies). These asset classes, particularly securities and commodities, remain part

of the ongoing debate on cryptocurrency classification today. Using a data science

approach, this study predicts the volatility of cryptocurrency prices using fiat

currencies (US dollar, Euro, Yen, etc.), securities (S&P 500, etc.), and commodities

(gold, etc.) over the past ten years and identifies patterns of behavior that might

suggest an asset class. Study 1a includes other economic indicators, such the

Consumer Price Index (CPI) and the U.S. Unemployment Rate, to further improve the

predictability of the models.

Unlike prior studies which were looking for predictive relationships between

certain market forces and cryptocurrency prices, Study 1a expands upon previous

research in three ways: First, this study evaluates the performance of six different

cryptocurrencies; second, this study analyzes ten years of data using advanced machine

learning models, including neural networks, to predict volatility in cryptocurrency

prices; and, third, this study utilizes a comprehensive set of economic variables to

predict this volatility.

Study 1b expands upon the methodology used in Study 1a to assess volatility in

stablecoin prices. Stablecoin volatility is assessed by predicting each coin’s “depegging”,

or what causes the coins to fall below its stated $1 valuation. Like Study 1a, Study 1b

relies on a comprehensive suite of economic variables to predict the depegging of three

stablecoins: Tether (USDT), U.S. Dollar Coin (USDC), and Binance Coin (BUSD).

Study 2 focuses on a different perspective on cryptocurrency performance – an

examination of internal factors rather than market forces. Using traditional artificial

neural networks and related statistical modeling techniques, this study evaluates 12

blockchain performance variables, or technical attributes that measure the efficiency of a

blockchain (hash rate, average block size, etc.), to ascertain if any of them (or groups of

them) have predictive relationships with a cryptocurrency’s price volatility. Study 2

focuses specifically on Bitcoin because 1) it has the most historical data, and 2) its

blockchain performance variables are the most accessible. Few studies to date have

attempted to evaluate these variables as a comprehensive set for predictive properties. As

the technology that underpins cryptocurrency performance is continuing to evolve, this

examination of predictor variables that are unrelated to market forces may be

illuminating.

1.2 Purpose of the Study

The purpose of this study is to examine the economic and technical factors that

may influence cryptocurrency price volatility. Prior studies have assessed economic

indicators in relation to cryptocurrency markets; others have investigated the effects of

technical architecture on blockchain performance. Neither area of study has been able to

identify the ways in which cryptocurrency markets are influenced by economic factors or

technology. This study is the first of its kind to assess cryptocurrency markets through the

lens of both economic and blockchain performance.

Studies 1a and 1b seek to broaden the understanding of cryptocurrency behavior

by evaluating the nine largest traditional cryptocurrencies in terms of market

capitalization in comparison to the behaviors of other asset classes (such as securities,

commodities, and fiat currencies) and other economic indicators (such as CPI and the

U.S. Unemployment Rate) over a ten-year period. These asset classes, particularly

securities and commodities, remain part of the ongoing debate on cryptocurrency

classification. By comparing cryptocurrency behavior to these asset classes, this study

informs conversations about appropriate classification of crypto assets. This study also

examines the economic behavior of stablecoins, a subcategory of cryptocurrencies that are

backed by another asset. Stablecoins have, to date, received little to no attention in

academic literature.

Study 2 suggests that the relationship between economic performance and

cryptocurrency prices depends on the technical architecture of the cryptocurrency’s

blockchain. Unlike other technical architectures, blockchain design is characterized by

the ways in which it achieves consensus. In other words, blockchain architecture is the

mechanism by which cryptocurrency investors commit and verify transactions on a

blockchain. Since achieving consensus relies on the behavior of individuals, it is

reasonable to expect that the intersecting relationships between numerous variables that

represent a blockchain’s architecture might collectively influence cryptocurrency price

volatility. Specifically, Study 2 looks at the predictive properties of blockchain

performance variables on Bitcoin price volatility.

1.3 Research Questions

Study 1a seeks to broaden the understanding of cryptocurrency volatility by

evaluating the six largest traditional cryptocurrencies in terms of market capitalization –

Bitcoin (BTC), Ethereum (ETH), Binance Coin (BNB), Ripple (XRP), Cardano (ADA),

and Dogecoin (DOGE). Study 1b expands upon Study 1a by exploring prince volatility

of the three largest stablecoins by market capitalization – Tether (USDT), USD Coin

(USDC), and Binance USD (BUSD). Together, these cryptocurrencies were evaluated in

comparison to the behaviors of the three major asset classes (securities, commodities, and

fiat currencies), to other economic indicators over a ten-year period, and to other

cryptocurrencies.

Study 1a addresses the following research questions:

Question 1a: How effective are advanced machine learning models in

predicting cryptocurrency price volatility?

Question 1b: Which variables are most significant in cryptocurrency

price volatility?

Study 1b addresses the following research question:

Question 2a: How effective are advanced machine learning models in

predicting the depegging of a stablecoin?

Question 2b: Which variables are most significant in predicting

stablecoin depegging?

Study 2 looks at the predictive properties of blockchain performance variables on

Bitcoin price volatility.

Question 3a: How effective are advanced machine learning models in

predicting price volatility of Bitcoin?

Question 3b: Which blockchain variables are most significant in

predicting Bitcoin price volatility?

1.4 Overview of Study Design

Data science harnesses the power of big data, characterized by its volume, variety,

and velocity, to advance the analytics about cryptocurrency behavior (Delen, 2021). This

approach helps to answer descriptive questions such as “what happened”, “why is it

happening”, and “how often does it happen” and predictive questions such as “what else

is most likely to happen” and “how else will it happen” (Delen, 2021).

The three studies followed the Cross-Industry Standard Process for Data Mining

(CRISP-DM) process to examine over 500,000 data points (Delen, 2021; Shearer, 2000).

Various models were developed that best suit the data and the research questions posed in

this study. Broadly speaking, predictive models were developed to address the research

questions. Studies 1a and 1b seek to understand whether cryptocurrencies price volatility

can be predicted by assessing economic variables. These questions are not attempting to

use these factors to predict cryptocurrency prices. Study 2 looks to see if Bitcoin price

volatility can be predicted by certain blockchain performance variables. Several

predictive models were used in both studies.

1.5 Significance of the Study

Studies to date have identified some potential correlations between the

performance of cryptocurrencies and other asset classes or economic indicators.

However, their results are limited to the study of one or two cryptocurrencies and few

other predictor variables. As such, their results are conflicting and suggest few reliable

conclusions. This study examines over 500,000 publicly available data points spanning a

10-year period to:

1. Evaluate a suite of cryptocurrencies, including stablecoins which have little

representation in the academic literature;

2. Simultaneously examine the cryptocurrency price volatility using

numerous economic indicators;

3. Simultaneously evaluate a suite of internal characteristics (blockchain

performance variables) of cryptocurrencies for any predictive properties;

and

4. Utilize a data science approach which enables broad and comprehensive

analysis of cryptocurrency behavior.

The academic literature on the behavior of cryptocurrencies, and particularly

how their behavior relates to other economic or technical indicators, is sparse. Given that

cryptocurrencies emerged for the first time in 2008, this represents a new field of study

and therefore, a greenfield opportunity to contribute to the literature on the behavioral

characteristics of cryptocurrencies.

1.6 Assumptions

While the two studies in this paper examine several predictor variables (asset

classes, economic indicators, and blockchain performance variables), the number of

potential predictor variables (even within the broad categories of variables selected for

this paper) are limitless. As such, the specific predictor variables selected for these

studies represent a natural inherent assumption. A discussion of some of the potential

variables that may be related to cryptocurrency price can be found in Chapter II, Other

Potential Influences on Cryptocurrencies.

The timing and potential impact of new or expanded regulatory oversite that

relates specifically to cryptocurrency activity is not captured in these studies. The

Securities and Exchanges Commission (SEC) and Commodity Futures Trading

Commission (CFTC) are still developing their positions on the asset classification of

cryptocurrencies. Further, the Office of the Comptroller of the Current (OCC) issued its

first interpretative letter for banks in 2020, while the Federal Deposit Insurance

Corporation (FDIC) and the Federal Reserve issued their letters in 2022. Another

assumption in this paper is that the impact of new or expanded regulatory oversite has not

yet been realized.

As discussed in this paper, the two largest cryptocurrencies, Bitcoin and

Ethereum, utilize proof-of-work consensus protocols to validate blockchain transactions.

However, in September 2022, Ethereum moved to a proof-of-stake consensus which is

different in many material respects from proof-of-work (Ethereum, 2022). Another major

assumption of this paper is that the studied variables pre-date this change in Ethereum’s

validation approach.

At the time of this paper’s publication, the volatility of cryptocurrency markets is

a well-observed phenomenon. A “crypto winter” is broadly defined as a period in which

most cryptocurrencies experience lower values (Chohan, 2022). While a several crypto

winters have been observed in the past, the record prices of cryptocurrencies in 2020

made the crypto winter experienced in 2022 particularly obvious (and painful, for some)

(Chohan, 2022). As the 10-year scope of this paper includes data only to June1, 2022, it

is unclear how the market downturn of the 2022 crypto winter may impact the efficacy of

the paper’s predictive models.

Finally, certain global events have been observed to influence cryptocurrency

prices. For example, when facing economic sanctions, countries sometimes turn to crypto

markets to allow for cross-border payments. With the ongoing political events between

Ukraine and Russia, there is discussion that economic sanctions on Russia will encourage

more Russian activity in crypto markets, thus inflating the price of cryptocurrencies.

Another example was the sharp increase in cryptocurrency activity observed during the

Covid-19 pandemic, a time when U.S. interest rates were low and stock market

performance was volatile (Katsiampa, Yarovaya, & Zięba, 2022). The analysis described

in Chapter III did not attempt to tease out the impact of these (or other) global events.

1.7 Definition of Terms

Asset Class – A group of investments that exhibit similar economic behaviors. For this

study, three asset classes are considered: 1) securities (such as the S&P 500 and

Dow Jones Industrial Average), 2) commodities (such as gold spot price and

crude oil spot price), and 3) fiat currencies (such as the U.S. dollar, Euro, and

Japanese Yen).

Artificial Neural Networks (ANN) – Conceptually, neural networks are analogous to

networks in the human brain that can learn from past experiences. Artificial

neural networks are developed via neural computing which employs pattern

recognition for machine learning and predictive modeling.

Block – A block represents a set of transactions that is stored and encrypted on a blockchain.

Blocks are linked together cryptographically using hash values.

Blockchain – The most well-known type of Distributed ledger technology (DLT),

blockchain, originated as the architecture supporting the Bitcoin

cryptocurrency (Nakamoto, 2008a). Several technical variables that are unique

to a blockchain influence its overall performance or efficiency. These include

hash rate, block size, and mempool size. A description of the blockchain

performance variables considered in this study can be found in Chapter II,

Blockchain.

Cryptocurrency – A type of digital asset. They can be broadly grouped into two categories:

1) Traditional cryptocurrencies such as Bitcoin and Ethereum, and 2)

Stablecoins such as Tether and US Dollar Coin. Bitcoin is often recognized as the

world’s first cryptocurrency. Created in 2008, Bitcoin was the first electronic cash to use

distributed leger technology (Nakamoto, 2008a; Popper, 2015). The use of a blockchain,

one type of distributed ledger technology, distinguishes cryptocurrencies from traditional

currencies and the other digital payments mechanisms that preceded it.

Cryptography – Mathematical algorithms used for encrypting data. There are three

major classifications of cryptography: 1) symmetric key encryption, 2)

asymmetric key encryption, and 3) hash functions).

Digital Asset – Anything (such as art, manuscripts, money, etc.) that can be stored digitally

and is considered to have intrinsic value.

Distributed Ledger – A distributed ledger can be thought of as a database that is shared

and synchronized real-time across multiple nodes (Nofer, Gomber, Hinz, &

Schiereck, 2017). Unlike a cloud-based database where multiple users can view

and edit one master database stored in a centralized location, a distributed ledger

must have the support from most nodes before a modification to the ledger can be

made (Sankar, Sindhu, & Sethumadhavan, 2017). Blockchain is one type of

distributed ledger.

Cross Industry Standard Process for Data Mining (CRISP-DM) – A framework that

guides data science research. Steps in this data preparation framework

include data consolidation, data cleaning, data transformation, and data

reduction

(Shearer, 2000).

Consensus Protocols – Algorithms that ensure all nodes on a blockchain agree on which

transactions are legitimate and are valid to add to the blockchain. The two most

used consensus protocols are: 1) Proof-of-Work (PoW), 2) Proof-of-Stake (PoS).

Crypto Winter – A period in which most cryptocurrencies experience lower values.

Data Science – A methodological approach to analytics that utilizes machine learning

algorithms to reveal the prediction accuracy of certain variables. Models include

Artificial Neurol Network (ANN), Decision Tree (DT), Recurrent Neural

Network (RNN), Long Short-Term Memory (LSTM), Bayseian Neural Network

(BNN), and Random Forest (RF).

Financial Services – The broad classification of businesses in the financial industry

that include and support banks, credit unions, insurance companies, wealth

management firms, credit card companies, etc.

Fork – Phenomenon that occurs when blocks at a blockchains are mined at conflicting

times or when a consensus cannot be reached. In this event, the blockchain

splits in a process known as “forking” (da Silva et al., 2019), which causes

inconsistencies on the blockchain network.

Hash Rate – The speed at which blocks are added to a blockchain used a proof-of-work

consensus protocol. Higher the hash rates mean that there are more miners

validating transactions, so it is also a representation of the total computational

power used by the blockchain.

Hash Value – A cryptographic term which represents the mathematical calculation of the

data in a block. If the data in the block changes, the hash will also change. In a

blockchain, the hash for one block is used to create the hash for the next block.

This means that if the data in one block changes, it will change the hash of that block as

well as the hashes of every other block in the ledger (Di Pierro, 2017).

KNIME Analytics Platform – Open-source, advanced analytics software that is used in data

science research.

Mining – One method of recording and verifying transactions on a proof-of-work

blockchain. To verify transactions, miners must solve a complex hashing

problem. The first miner to find a solution is rewarded by receiving newly

minted crypto coins from that specific blockchain.

Proof-of-Stake (PoS) – A consensus protocol which chooses the creator of a new block

based on the wealth, or stake, of the verification node. PoS blockchains do not

provide a block reward. Instead, the nodes performing validation, called forgers,

charge fees for each transaction on the blockchain. These fees, also called “gas

prices”, fluctuate with demand. Unlike PoW, complex mathematics is not used to

validate transactions, so the protocol requires virtual no consumption of energy.

Ethereum 2.0 converted from PoS to PoW in September 2022.

Proof-of-Work (PoW) – Requires that a subset of nodes, called miners, calculate a

mathematical problem, called the “proof-of-work”, to determine who can validate

transactions on the blockchain and receive a block reward (Gervais et al., 2016).

In recent years, proof-of-work consensus has come under scrutiny due to its heavy

consumption of energy and other limited resources (Xue et al., 2018). PoW is the

consensus protocol used by Bitcoin.

Stablecoin – A type of cryptocurrency that attempts to peg its value to another financial

instrument. For example, the values of Tether and Binance USD are supposed to

be pegged to the value of the U.S. dollar. In theory, stablecoins were designed to

offer price stability and less volatility. In practice, however, many stablecoins

have struggled to achieve this objective.

Volatility or Daily Volatility – The statistical measure of the dispersion of return, for Bitcoin.

Daily volatility was calculated using the Garman-Klass (GK) volatility measure

(Garman & Klass, 1980).

CHAPTER II

II. CONCEPTUAL BACKGROUND AND REVIEW OF LITERATURE

2.1 Introduction

Money, whether it is represented by a stone, forged coin, paper note, or line of

computer code, does not implicitly have intrinsic value. The value of money is dependent

on the perception of its importance by the people who use it as a medium of exchange, a

store of value, or a unit of account. While there have been countless developments in

monetary technology over the course of human history, cryptocurrencies represent a

novel monetary development, both technically and economically. Technically,

cryptocurrencies are decentralized, meaning that no single entity controls the information

or infrastructure. Economically, cryptocurrencies exist digitally and, as a result, appear to

have different interactions with the money of the physical world. Together, these two

attributes make comparisons between cryptocurrencies and traditional financial assets

worthy of exploration.

The academic literature on cryptocurrencies has primarily focused on the pricing

and valuation of cryptocurrencies (Alessandretti, ElBahrawy, Aiello, & Baronchelli,

2018; Sovbetov, 2018). Research on the behavior of cryptocurrencies, and particularly

how their behavior relates to other financial asset classes, is sparse. Given the steady

increase in cryptocurrency activity over the past decade, and the sharp increase during the

Covid-19 pandemic, this represents a new field of study and a greenfield opportunity to

contribute to the literature on the behavioral characteristics of cryptocurrencies.

2.2 Longitudinal Review of Cryptocurrency Development

2.2.1 History of Cryptocurrencies

While it is commonly believed that cryptocurrencies emerged in 2008, the earliest

attempt at distributed transactions can be traced back to a company called DigiCash

founded by David Chaum in 1989 (Nakamoto, 2008a; O'mahony, Peirce, & Tewari,

1997). Chaum created the concept of blind signatures which allowed for anonymity in

digital transactions (Chaum, 1983). The prefix “crypto” is derived from the Greek word

κρυπτός meaning concealed, secret, or hidden. The ability to create anonymity in

transactions was essential to early cryptocurrency research. Blind signatures are a

derivation of digital signatures which rely on public-key cryptographic hashes to validate

the integrity and authenticity of a digital transaction, message, or document (Diffie &

Hellman, 1976; Merkle, 1987). DigiCash used blind signatures to disguise transactions

before posting to the network thus allowing for anonymous transactions on a public

ledger (Chaum, Fiat, & Naor, 1988). Blockchain technology would come to rely on blind

signatures to facilitate anonymous transactions (Zheng, Xie, Dai, Chen, & Wang, 2018).

In 1993, computer scientists from Bell Research Laboratory and IBM worked alongside

the Financial Services Technology Consortium (FSTC), a consortium of banks,

technology companies, and government entities, to develop a hash-based system for

protecting blocks of data within a newly design electronic check processing format

(O'mahony et al., 1997) (see Appendix I for FSTC members in 1998). To facilitate the

development of uniform structural components of an electronic check., the FSTC

developed the Financial Service Market Language (FSML) which included the creation

of blocks to represent components of physical checks (O'mahony et al., 1997). According

to Donal O’mahony, the block structure was set up such that “[a] signature block

contains the names of the other blocks that is signs, along with a hash of each of those

blocks. Therefore, the e-check signature is actually a signed hash of the concatenated

hashes of several blocks.” (O'mahony et al., 1997, p. 142).

Daniel Schutzer, former Vice President and Director of External Organizations,

Standards, and Advanced Technology at Citibank as well as former President of the

FSTC discussed the need for “electronic money” in Banking and Finance on the

Internet, published in 1998, describing it as a departure from paper currency (Schutzer,

1998). He goes on to elaborate on benefits of electronic cash stating: “[t]here are strong

incentives for moving to electronic money, including increased convenience, timeliness,

flexibility, and control of payments both on and off the ‘Net...[it] requires some changes

in how consumers think about money, and a total shift will not occur overnight.”

(Schutzer, 1998, p. 183). The work of the FSTC represents an important, and often

overlooked, element in the development of modern cryptocurrencies. For example,

Schutzer directly references earlier iterations of digital cash, including DigiCash

(Schutzer, 1998, p. 189).

He also describes a future state initiative to develop “electronic wallets” to host digital

cash (Schutzer, 1998, p. 202). Many scholars acknowledge cryptocurrency research and

development by entrepreneurs and computer scientists, but few recognize the work done

by financial services organizations. Despite widespread recognition, the FSTC appears to

be directly involved in the development of decentralized ledgers and digital currencies.

In 2002, Adam Back released a white paper which described an email filtration process

based on a proof-of-work system (Back, 2002). The email filtration process relied on a

cost-function, originally developed by Cynthia Dwork and Moni Naor, which required the

sender to calculate a computationally expensive problem to send a message (Back, 2002;

Dwork & Naor, 1992). The proposed proof-of-work model required clients to compute

tokens (𝒯) using cost-function MINT (𝑠,𝑤), where w is the amount of work required

to mint a token, to participate in a protocol with a server in processes termed minting. The

server verifies the value of the token using the function VALUE (). This process reduces

the likelihood a server would experience resource depletion as is commonly the case with

Denial of Service (or DOS) attacks (Back, 2002; Moore, Shannon, Brown, Voelker, &

Savage, 2006). Nakamoto used the idea of a proof-of-work to develop the consensus

protocol that powers the Bitcoin blockchain (Back et al., 2014; Nakamoto, 2008a).

Other efforts focused on the development of decentralized digital currency. In

1998, computer engineer Wei Dai published the white paper “b-money, an anonymous,

distributed electronic cash system” where Dai defines the b-money protocol, a protocol

that eventually served as a blueprint for Bitcoin (Dai’s paper is directly cited in the

Bitcoin white paper) (Dai, 1998; Nakamoto, 2008a). The b-money protocol is a set of

two protocols. The first protocol relies on a broadcast channel in which every participant

keeps and maintains a copy of the database tracking the amount of money owned by

given individuals. This protocol also describes a money creation system in which users

solve computational problems that “have no value, either practical or intellectual” in

order to receive newly minted coins (Dai, 1998). The money creation process described

by Dai is the first introduction to the cryptocurrency creation process that will become

known as “mining” (Bhaskar & Chuen, 2015). At the time of publication in 1997, Dai

described the first protocol as “impractical” because of its heavy reliance on a

decentralized anonymous ledger system. The second b-money protocol reduces the

reliance on a completely decentralized ledger system by assigning transaction

verification responsibilities to a subset of participants (Dai, 1998). In this protocol, Dai

explains that the verification participants must hold a certain deposit of the virtual

currency to use in case of fines or rewards. The second protocol’s reliance on a subset of

participants for transaction verification would eventually be incorporated into the Bitcoin

mining process (Bhaskar & Chuen, 2015).

Development on electronic currencies, or underlying distributed ledgers, appears to slow

rapidly in the early to mid-2000s. Then, in the midst of the Great Recession on

October 31, 2008, Satoshi Nakamoto sends a whitepaper titled Bitcoin: A Peer-to-Peer

Electronic Cash System to an online cryptography mailing group called Metzdowd

(Nakamoto, 2008b). The paper is short, only eight pages in length, and includes

eight references, some of which specifically reference b-money and Hashcash

(Nakamoto, 2008a). The paper lays the foundation for what will eventually become

Bitcoin, the world’s first cryptocurrency (Popper, 2015). Its author, Satoshi

Nakamoto, was determined to be a pseudonym whose true identity has never been

revealed (Popper,

2015).

Following publication of the Bitcoin whitepaper, cryptocurrencies exploded as a

market force with the potential for applications ranging from payments to central bank

digital currencies (CBDCs). In 2014, Vitalik Buterin published Ethereum White Paper: A

Next Generation Smart Contract & Decentralized Application Platform to Metzdowd, the

same cryptography mailing list where Satoshi Nakamoto published the Bitcoin

whitepaper (Buterin, 2014). Then in 2019, Facebook announced their development of a

stablecoin, Libra, via their consortium, the Libra Association. Libra never launched and

in 2021 the remaining intellectual property was sold to a private bank (Arauz, 2021;

Kharif, 2022; The Libra Association, 2019). Since then, other consortia have since

emerged to develop cryptocurrencies which are still underway (Phillips, 2022). Despite

these advancements in cryptocurrency development, debate remains as to whether

cryptocurrencies are money, securities, commodities, or their own thing entirely.

2.2.2 Types of Cryptocurrencies

Traditional cryptocurrencies. Popular cryptocurrencies, such as Bitcoin or

Ethereum, are considered “traditional” cryptocurrencies as they rely on the basic

underlying structure and assumptions put forward by Satoshi Nakamoto. These

cryptocurrencies are exchanged on public blockchains where transaction information

is openly available and do not have any backing assets, meaning their value is derived

entirely by market forces (Ante, Fiedler, & Strehle, 2021). While traditional

cryptocurrencies remain popular, their high price volatility is believed to prevent them

from being used as a mainstream means of payment (Manaa et al., 2019).

Stablecoins. Unlike traditional cryptocurrencies, stablecoins are designed to maintain

a stable value relative to other assets, typically national fiat currencies (President's

Working Group on Financial Markets, Federal Deposit Insurance Corporation, &

Office of the Comptroller of the Currency, 2021). Stablecoins have expanded use cases

compared to traditional cryptocurrencies – stablecoins are used to facilitate trading and

payments as well as lending and borrowing other cryptocurrencies (President's

Working Group on Financial Markets et al., 2021). Popular stablecoins include Tether,

U.S. Dollar Coin (USDC), and Binance USD.

Other cryptocurrencies. Many other types of cryptocurrencies exist with new

variations constantly entering the marketplace. One specific type of cryptocurrency

worthy of note is central bank digital currencies (CBDCs). Several nation states, and their

corresponding central banks, are interested in creating a national cryptocurrency. They

often utilize many of the properties of stablecoins to create these national

cryptocurrencies, referred to as CBDCs. Countries currently researching, testing, or

offering a CBDC include: the Bahamas (Central Bank of The Bahamas, 2019), the

European Union (European Central Bank, 2020), China (People's Bank of China, 2021),

England (Bank of England, 2021), and the United States (Federal Reserve Bank of

Boston & Massachusetts Institute of Technology Digital Currency Initiative, 2022).

2.2.3 Overview of Cryptocurrency Prices

By design, cryptocurrencies are decentralized assets. They do not require

intermediation from banks and therefore can largely evade control by governments and

central banks. Without connection to central banks, cryptocurrencies have no inherent

connection to the ‘real’ economy. For this reason, it is reasonable to expect that

cryptocurrencies could be used to hedge against uncertainty in traditional markets such as

the stock market or traditional fiat currencies. Baek and Elbeck (2015) demonstrated this

point by showing that fluctuations in cryptocurrency prices are driven by market

participants. Even more interesting, Klein, Thu, and Walther (2018) showed that Bitcoin

is positively correlated with downward markets suggesting that cryptocurrencies serve as

a store of value during declines in the market (Conlon & McGee, 2020).

2.3 Predicting Cryptocurrency Price Volatility via Existing Asset Class and Economic Indicator

Data

Studies 1a and 1b seek to understand whether the volatility of cryptocurrency

prices can be predicted by other asset classes or economic indicators. The studies

evaluate cryptocurrencies, including “traditional” cryptocurrencies and stablecoins,

commodities, securities, fiat currencies, and other economic indicators.

2.3.1 Cryptocurrencies as Currencies

Given their name, cryptocurrencies are often mentioned in colloquial

conversation as if they are currencies, like the yen or euro. However, governments,

economists, and computer scientists disagree on the appropriate classification for

cryptocurrencies. In the United States, the Internal Revenue Service (IRS) issued notice

2014-21 stating that cryptocurrencies are to be treated as property, not currencies

(Internal Revenue Service, 2014; Lerer, 2019). Economists, on the other hand, are not

consistent in their view of cryptocurrencies. Kirkby (2018) found similarities in the

performance of Bitcoin and the Venezuelan bolívar but concludes that cryptocurrencies

would “be a bad idea” (p. 528).

(White, 2015) describes cryptocurrencies as “competing private irredeemable

monies” but Smith and Kumar (2018) clarify that cryptocurrencies have a lack of

institutional liability and, therefore, cannot be deemed irredeemable monies such as

bank deposits.

Economists identify three characteristics of money: it serves as a store of value, a

unit of account, and a medium of exchange (Hennerich, 2021). Many researchers agree

that various assets, including cryptocurrencies, can be used as units of account; likewise

there is consensus that in order to become a medium of exchange, an asset must meet

the other two qualities (Ammous, 2018). This leaves store of value as the property to

scrutinize in the context of cryptocurrencies. Unlike fiat currencies, cryptocurrencies

have no central banks, and, therefore, no mechanism to set interest rates and control

money supply growth. Instead, cryptocurrency supply growth is set by predetermined

algorithms that not only serve as economic policy, but also finance the infrastructure

required to support transactions. Ammous (2018) found that Bitcoin has had an annual

growth rate over 100% since 2013, with growth rates expected to increase to 200% by

2024. By comparison, broad money growth for 167 countries between 1960 and 2015

averaged 32.16% with the money supply growth for Organization for Economic

Cooperation and Development (OECD) countries dropping to an average rate of 7.17%

from 1990 to 2015 (Ammous, 2018). This suggests that the growth rate of Bitcoin is too

high to allow for it to become a store of value.

2.3.2 Cryptocurrencies as Securities

For those who view cryptocurrencies as assets, the next question becomes: what type

of asset? Gary Gensler, current chairman of the U.S. Securities and Exchange

Commission (SEC), stated that he believes “the vast majority of [cryptocurrencies] are

securities” (Gensler, 2022). Gensler’s rationale for this belief is that cryptocurrency

investors are expecting to earn a profit making cryptocurrencies investment contracts

under the U.S. Supreme Court’s 1946 Howey Test (Gensler, 2022; Henning, 2018).

Gensler’s predecessor, former SEC Chairman Jay Clayton, told journalists from CNBC

that “cryptocurrencies are replacements for sovereign currencies, replace the dollar, the

euro, the yen with bitcoin. That type of currency is not a security” (Rooney, 2019).

Researchers, like SEC chairmen, are unclear if cryptocurrencies function as securities.

Bitcoin, for instance, appears to not act safe-haven for the S&P 500 (Conlon & McGee,

2020). However, cryptocurrencies issued via initial coin offerings (ICO), the act of raising

money by issuing cryptocurrency tokens, appear to behave similarly to public equity

offered by initial public offerings (IPOs) (Lyandres, Palazzo, & Rabetti, 2019). This

implies that that ICO tokens do behave like securities. Further analysis is needed to

understand the behavior pattern of cryptocurrencies relative to securities.

2.3.3 Cryptocurrencies as Commodities

If not securities or currencies, then perhaps cryptocurrencies can be seen as

commodities. The U.S. Commodity Futures Trading Commission (CFTC) states that

cryptocurrencies, like Bitcoin, meet the definition of a commodity under the Commodity

Exchange Act (U.S. Commodity Futures Trading Commission, 2020). However, like

securities and currencies, research is not clear on the performance of cryptocurrencies as

commodities. Klein et al. (2018) found that Bitcoin does not serve as a safe-haven

investment, the prominent feature of gold, and therefore does not consider Bitcoin to be a

commodity (often referred to as the “new gold”). Other studies show that

cryptocurrencies like Bitcoin and Ethereum serve as a functional hedge and safe-haven

for individual commodities (Naeem, Farid, Balli, & Hussain Shahzad, 2021; Okorie &

Lin, 2020). Significant discrepancies exist in the literature as to whether cryptocurrencies

behave, and should be considered, commodities.

2.3.4 Other Economic Influences on Cryptocurrency Prices

Some studies suggest that the only statistically significant indicator of cryptocurrency

performance is the spread between daily high and low prices (Baek & Elbeck, 2015). In

other words, cryptocurrency prices cannot be reliably predicted using economic

modelling for traditional assets. This suggests that cryptocurrencies may be an entirely

new asset class.

2.4 Predicting Bitcoin Price Volatility Using Blockchain Performance Variables

Study 2 looks at the predictive properties of certain internal blockchain

performance variables on Bitcoin price volatility given that the technology that underpins

cryptocurrency performance continues to evolve.

2.4.1 Overview of Blockchain

Blockchain technology originated as the ledger system supporting the

cryptocurrency Bitcoin (Nakamoto, 2008a). A distributed ledger can be thought of as a

database that is shared and synchronized real-time across multiple nodes (Nofer et al.,

2017). Unlike a cloud-based database where multiple users can view and edit one master

database stored in a centralized location, a distributed ledger must have the support from

most nodes before a modification to the ledger can be made (Sankar et al., 2017).

Each block of transactions on a distributed ledger has a unique identifier, known as a

hash. Hash values are mathematical calculations based on the data in the block itself, so if

the data in the block changes, the hash will also change (Di Pierro, 2017). In a

blockchain, the hash for one block is used to create the hash for the next block. This

means that if the data in one block changes, it will change the hash of that block as well

as the hashes of every other block in the ledger (Di Pierro, 2017). The collection of

interconnected blocks is why this technology is called “blockchain” (Figure 1).

Figure 1. Graphical Depiction of Blockchain Technology

Since a copy of the distributed ledger, or blockchain, is maintained on each node

in the network, the need for a central authority is eliminated (Pelt, Jansen, Baars, &

Overbeek, 2021). The decentralized nature of a blockchain also makes it more resilient,

though not immune, to traditional cyber-attacks since each copy of the ledger would need

to be attacked simultaneously for the attack on the blockchain to be successful (Sayeed &

Marco-Gisbert, 2019). Because there is no single holder of the information stored on a

blockchain, as is the case with traditional databases, blockchains are tamper-resistant

while also serving as a distributed source of truth for all transactions (Li, Jiang, Chen,

Luo, & Wen, 2020).

2.4.2 Consensus Protocols

Consensus protocols ensure all nodes on a blockchain agree on which transactions are

legitimate and valid to add to the blockchain. Consensus algorithms establish the

protocol rules and ensure compliance with the rules to ensure that all transactions occur

in a trustless manner. Consensus protocols facilitate agreement, collaboration,

cooperation, equal rights for each node, and mandatory participation of each node in the

consensus process.

The technical origin of consensus protocols is the theory of “Byzantine fault

tolerance” which describes a situation in which a computer system, particularly

distributed computer systems, fail without clear information that the system has failed

(Castro & Liskov, 2002). In blockchain, the establishment of a Byzantine fault tolerance

allows trustless networks to function even when malicious nodes are present (Gramoli,

2020). Consensus protocols establish the Byzantine fault tolerance for blockchain-based

systems.

There are many different types of consensus protocols. The most popular protocol, and

the protocol used by many cryptocurrency blockchains including Bitcoin, is proof-of-

work (PoW). PoW consensus requires that a subset of nodes, called miners, calculate a

mathematical problem, called the “proof-of-work”, to determine who can validate

transactions on the blockchain and receive the block reward (Gervais et al., 2016). In

recent years, proof-of-work consensus has come under scrutiny due to the heavy

consumption of computing power, energy, and other limited resources (Xue et al.,

2018).

Proof-of-stake (PoS) is another way to validate transactions and achieve

distributed consensus. Unlike the PoW, PoS chooses the creator of a new block based on

the wealth, or stake, of the verification node. PoS blockchains do not provide a block

reward. Instead, the nodes performing validation, called forgers, charge fees for each

transaction on the blockchain. These fees, also called “gas prices”, fluctuate with demand.

In delegated proof-of-stake (DPoS), nodes on the blockchain vote for witnesses to secure

the network. The vote strength of a particular node is determined by the size of their

stake. Nodes with a larger stake will influence the network more than nodes with a lesser

stake. Version 2.0 of the Ethereum blockchain, deployed in September 2022, uses a

PoS consensus protocol (Ethereum, 2022).

Other consensus protocols include proof-of-elapsed time (PoET) (Castro &

Liskov, 1999), proof-of-benefit (Liu, Liu, Jiang, Zhao, & Wang, 2018), proof-of-learning

(Bravo-Marquez, Reeves, & Ugarte, 2019), and proof-of-accuracy (Kudin, Kovalenko, &

Shvidchenko, 2019). While these consensus protocols are being researched for their

applicability to cryptocurrencies, they are not widely used at this time (Oyinloye, Teh,

Jamil, & Alawida, 2021).

2.4.3 Mining

Blockchain maintenance, commonly referred to as mining on proof-of-work

blockchains, involves the recording and verifying transactions. To verify transactions,

miners must solve a complex hashing problem. The first miner to find a solution is

rewarded by receiving newly minted crypto coins from that specific blockchain (Lo &

Medda, 2019). There are two types of mining: solo mining and pool mining. In solo

mining, each miner uses their personal hardware and registers themselves as a miner.

The first miner to find the solution informs the other miners that the solution has been

found. In solo mining, the rewards are not shared (Qin, Yuan, & Wang, 2018). By

contrast, in pool mining, individuals do not have enough computing power to mine

resources on their own. Instead, the miners combine resources to mine the blockchain

faster. Pool mining allows for economies of scale since the cost of mining is lower for all

miners. Although the mining rewards are shared in pool mining, there is greater

opportunity for higher mining income (Salimitari, Chatterjee, Yuksel, & Pasiliao, 2017).

2.4.4 Bitcoin Price Volatility and the Impact of Blockchain Performance

While some scholars, such as (Buchholz, Delaney, Warren, & Parker, 2012), find

that changes in Bitcoin price (as price of any currency) is directly impacted by supply

and demand, there is a growing body of research that suggests the price volatility of

Bitcoin cannot be explained using traditional economic and financial theories, such as

future purchasing power parity, uncovered interest rate parity, or cash-flow models,

because supply and demand fundamentals do not apply to Bitcoin markets. (Ciaian,

Rajcaniova, & Kancs, 2016; Gordon, 1962; Kristoufek, 2013; Krugman & Obstfeld,

2009; Levi, 2005).

The literature (Table 1) shows conflicting results on the effect of blockchain variables

on Bitcoin price.

Table 1. Review of the Literature on Blockchain Variables and Bitcoin Price

Authors Models Variables Considered Results

Chen

(2023)

Random forest,

LSTM

47 variables including

Bitcoin price, total fees, hash

rate, 10-year Treasury Yield

Bitcoin price

after October

2018 has become

less predictable;

U.S. stock

market index, the

price of oil, ETH

price, and the

difficulty of

finding blocks of

Bitcoin impact

price.

Ji, Kim,

and Im

(2019)

DNNs, LTSM,

CNNs,

ResNets, and

ensemble

models

Bitstamp Bitcoin market

(USD) time series data, 29

variables on bitcoin

blockchain including hash

rate, average block size,

and

Predictability of

all deep learning

models studied in

this work was

comparable.

number of transactions per

block

Aalborg,

Molnár,

and de

Vries

(2019)

Predictive

regression

Return, volatility, trading

volume, transaction volume,

change in the number of

unique Bitcoin addresses, the

VIX index and Google

searches for “Bitcoin”

Considered

variables cannot

predict Bitcoin

returns.

Guizani

and Nafti

(2019)

Auto

Regressive

Distributed

Lag ARDL

model,

cointegration

test (Pesaran,

Shin, & Smith,

2001),

Granger

causation test

as defined by

Toda and

Yamamoto

(1995)

Number of addresses,

attractiveness indicator,

mining difficulty, transaction

volume, stock prices,

EUR/USD exchange rate,

macroeconomic

development

Number of

addresses,

attractiveness

indicator, and

mining difficulty

have

significant

impact on BTC

price.

McNally,

Roche,

and Caton

(2018)

RNN, LSTM,

ARIMA

Bitcoin price index (BPI),

the closing price, the

opening price, daily high

and daily low are also

included as well as

blockchain data such as the

mining difficulty and hash

rate

RNN and LSTM

are effective for

Bitcoin

prediction.

Estrada

(2017)

Granger

causation test

as defined by

Toda and

Yamamoto

(1995)

Bitcoin price and the S&P

500, Bitcoin price and the

VIX, Bitcoin

realized volatility and the

S&P 500, and Bitcoin

realized volatility and the

VIX

No statistically

significant

Granger

causality when

analyzing the

time series of

BTC and the

S&P 500 and

BTC and the

VIX.

Ciaian et al.

(2016)

Augmented

Dickey-Fuller

test, the Phillips-

Perron test,

Granger

causation test,

vector error

correction

model

Bitcoin price, total number

of Bitcoins in circulation,

total unique Bitcoin

transactions per day, number

of unique Bitcoin addresses

used per day, Bitcoin days

destroyed for any

transaction, exchange rate

between U.S. dollar and

Euro, oil price, and Dow

Jones stock market index

Supply demand

fundamentals

have a strong

impact on

Bitcoin price.

The Dow Jones

Index, exchange

rate and oil

price

do not

significantly

affect Bitcoin

price.

Kristoufek

(2015)

Wavelet

analysis,

specifically

wavelet

coherence

Bitcoin price index (BPI),

total Bitcoin in circulation,

number of transactions,

estimated output volume,

trade volume vs. transaction

volume ratio, hash rate,

difficulty, exchanges, search

engine trends, Financial

Stress Index (FSI), gold

price,

Usage in trade,

money supply

and price level

play a role in

Bitcoin price

over the long

term.

Van Wijk

(2013)

Ordinary Least

Squares

regression,

vector error

correction

model

Closing value of the

exchange rate of Bitcoin,

closing value of the Dow

Jones Index, closing value of

the FTSE 100 Index, closing

value of the Nikkei 225

Index, exchange rates

between the U.S. dollar and

the Euro and the Yen, Brent

oil price, West Texas

Intermediate (WTI) oil price,

UBS Bloomberg

Constant Maturity

Commodity Index (CMCI)

of Oil

The Dow Jones

index, the

eurodollar

exchange rate,

and oil price

significantly

impact Bitcoin

prices.

Buchholz

et al.

(2012)

Augmented

Dickey Fuller

Tests, Breusch-

Godfrey test,

Supply of Bitcoin in

existence, total number of

Bitcoin transactions per day,

total value of Bitcoin

Volatility

significantly

effects price,

significantly

Engle-Granger transactions, price of Bitcoin less volatility

from

2.4.5 Blockchain Performance Variables – Not Previously Studied

Study 2 includes some additional blockchain performance variables related to

Bitcoin that have not received attention in the academic literature.

Average Block Size – The size of a block on a blockchain represents the amount of

data that the block can store. A year after Satoshi Nakamoto mined the first

Bitcoin genesis block, he modified the protocol to limit the size of these blocks to

1 MB (which can store 200-500 transactions). This decision was controversial

because, within certain limits, larger blocks enable higher transaction-per-second

rates. By 2017, a variety of protocol changes enabled much larger blocks on

Bitcoin’s blockchain.

Hash Rate – Hash rates reflect the amount of computing power on a network and

are computed by the number of transactions that can be performed on a network

per second. Relative to cryptocurrency mining, higher hash rates typically mean

that there are more miners on that network. They are also a measure of both the

health and security of a cryptocurrency network.

Consensus Difficulty – As previously described, Bitcoin uses a Proof-of-Work

(PoW) consensus protocol (decentralized and fault-tolerant) to validate

transactions on its network. PoW has shown to be relatively resilient and secure,

however it requires significant computing power and can be slow when networks

are busy. Consensus difficulty is a relative measure of how difficult it is to mine

test, vector

error

correction

model, ARCH

and GARCH

models

in U.S. dollars, average

transaction value, Google

searches, daily mentions

“bitcoin” on Twitter, daily

mentions of “Bitcoin” on

RSS-News feed.

negative shocks

compared to

positive shocks,

volatility

significantly

effects demand

a new block based on how much hashing power has been deployed on the

network.

(Mingxiao, Xiaofeng, Zhe, Xiangwei, & Qijun, 2017).

Transactions Per Block – A new Bitcoin block is generated every 10 minutes or

so. These blocks contain a record of every transaction and all those that occurred

before them (often between 200-500 transactions in each block). This variable

measures the average number of transactions per block in the last 24 hours.

Mempool Size – The number of transactions waiting to be confirmed by miners.

It is also a reflection of congestion on a network: the higher the mempool size, the

longer the transactions time might be.

Total Transaction Fees (USD) – This measure represents the total USD value of all

transaction fees paid to miners in a 24-hour period.

Miners’ Revenue (USD) – The dollar amount includes the transaction fees paid to

miners and the value of the Bitcoin blocks received as a reward for processing the

transaction confirmations.

Cost Per Transaction – This is calculated by the miners’ revenue divided by the number

of transactions.

Median Confirmation Time – This is the median time for a transaction with miner

fees to be included in a mined block and added to the public ledger.

2.5 Other Potential Influences on Cryptocurrencies

While the studies in this paper examine several predictor variables (asset classes,

economic indicators, and blockchain performance variables), the number of potential

predictor variables (even within the broad categories of variables selected for this

paper) are limitless. Forking is one of the variables not specifically considered in this

paper but is worthy of discussion and represents an opportunity for future research.

When blocks on a blockchain are mined at conflicting times, or when a consensus

cannot be reached, the blockchain splits in a process known as “forking” (da Silva et

al., 2019). Forking causes inconsistencies on the blockchain network. When a fork

happens and transactions are mined on both ‘sides’ of the forked blockchain, those

transactions on the side that is ultimately not accepted must be mined again. Studies

show that forking can impact of the performance of a blockchain (da Silva et al., 2019).

If we assume that blockchain performance is a critical component to cryptocurrency

prices, then it is reasonable to expect that activities that reduce blockchain performance

may reduce the value of cryptocurrencies as well.

CHAPTER III

III. METHODOLOGY

3.1 Introduction

Data science harnesses the power of big data, characterized by its volume, variety,

and velocity, to advance the analytics about cryptocurrency behavior (Delen, 2021). This

approach helps to answer descriptive questions such as “what happened”, “why is it

happening”, and “how often does it happen” and predictive questions such as “what else

is most likely to happen” and “how else will it happen” (Delen, 2021).

3.2 Approach

The studies in this paper follow the Cross-Industry Standard Process for Data Mining

(CRISP-DM) process to examine over 500,000 data points (Delen, 2021; Shearer,

2000). CRISP-DM provides a 6-step approach to data mining:

1) Business Understanding – In this step, an understanding of the business

circumstances and related problem(s) is developed. Chapter II outlines both

the historical and current state of cryptocurrency markets and suggests several

questions for which data mining might be applicable.

2) Data Understanding – The available data (all of which is public for this study)

is matched to the business problem. This chapter outlines the study’s data

sources and their applicability to the research questions. Research studies in

this space tend to employ one of two different strategies depending on the

types of data they evaluate. Structured data, or that which is numerical or

categorical, is best analyzed through derivations of regressions and neural

networks. Unstructured data such as text, images, and video is better served by

approaches that use natural language processing methodologies like sentiment

analysis (Delen, 2021), and there are some interesting recent studies of

cryptocurrency behavior that have used this approach (Karalevicius,

Degrande, & De Weerdt, 2018; McMillan, Myers, Nguyen, Robinson, &

Kennard, 2022; Polasik, Piotrowska, Wisniewski, Kotkowski, & Lightfoot,

2015). Given that all the variables to be evaluated in this study are numerical,

only structured data analytics techniques were used.

3) Data Preparation – Prior to mining, raw data was prepared (preprocessed) to

improve its quality and avoid the potential for errors due to inconsistencies,

erroneous values, or other GIGO (garbage-in/garbage-out) problems. See

Figure 2 below.

4) Model Building – In this step, various models were developed that best suit

the data and the research questions posed in this study. Studies 1a and 1b

(questions 1a, 1b, 2a, and 2b) seek to understand whether the price volatility

of cryptocurrencies aligns with that of certain asset classes or economic

indicators. Study 2 (questions 3a and 3b) looks to see if Bitcoin’s price

volatility can be predicted by certain blockchain performance variables. For

all these questions, the predictive models described in the Testing and

Evaluation section below were used.

5) Testing and Evaluation – Once the models were built, this step evaluated the

findings for accuracy and assesses the findings’ applicability to the business

problem.

6) Deployment – In this final step, the findings are presented in such a way they

are accessible to a variety of end-users.

The open-source advanced analytics tool, KNIME Analytics Platform 4.7.2, was

used for data preparation, model development, testing, and evaluation. KNIME Analytics

Platform was selected because of its open and expandable architecture which combines

the functionality of several tools (including Excel, Power BI, and Python) into a single

workflow (Delen, 2021).

3.3 Data Set and Procedures

Publicly available data from a variety of sources (see Tables 2-6) over the ten

years spanning January 1, 2012, through June 1, 2022, was mined through a variety of

methodological approaches. Prior to mining, raw data was prepared (preprocessed) to

improve its quality and avoid the potential for errors due to inconsistencies, erroneous

values, or other GIGO (garbage-in/garbage-out) problems. Data preparation occurred in

multiple phases (

Figure 2): data consolidation, data cleaning, data transformation, and data reduction

(Delen, 2021).

Figure 2. Process for Data Preparation

3.4 Measures

Six types of variables were examined to inform the research questions posed in

the two studies: Cryptocurrency prices, security prices, commodity prices, fiat currency

prices, economic indicator values or prices, and blockchain performance variables. These

variables were chosen specifically because they are broad, global, well-understood, and

represent the most likely factors to predict cryptocurrency volatility. Further, these

variables, particularly securities and commodities, continue to be part of the ongoing

debate on cryptocurrency classification. The variables are quantitative, and specifically,

they are either date stamps, prices, or numerical values. They are available through

reliable public sources dating back to 2012 and with a granularity (frequency) that lends

itself to data science methodologies. In fact, this study examines over 500,000 data points

to explore its research questions. The Data Dictionary in Appendix II details all of the

variables and descriptions; however, they are summarized as follows:

3.4.1 Dependent Variables

Cryptocurrency Price. All the research questions posed in this study are focused

on the volatility of cryptocurrencies as measured by price. Studies 1a and 1b considered

six cryptocurrencies and three stablecoins, respectively (Table 2Table 2). Stablecoins are

a type of cryptocurrency that are designed to peg their values to other financial

instruments. For example, the values of Tether and Binance USD are supposed to be

pegged to the value of the U.S. dollar. Study 2 also looked at price volatility but only for

Bitcoin.

Prices for the cryptocurrency in these studies are all publicly available and were

downloaded from nasdaq.com. All of these cryptocurrencies and stablecoins were also

included as predictor variables. Additionally, Solona (SOL) which was excluded as a

dependent variable because of its limited history, was included as a predictor variable in

studies 1a and 1b.

Table 2. Cryptocurrencies Considered in the Studies

Bitcoin (BTC) $557b Traditional The world’s first cryptocurrency

and the largest cryptocurrency by

market capitalization. Capped total

coin issuance at 21 million coins.

Ethereum (ETH) $219b Traditional The second largest

cryptocurrency by market

capitalization and the platform

that allows for smart contracts.

Tether (USDT) $72b Stablecoin Tether is the largest stablecoin with

its value pegged to the U.S. dollar.

U.S. Dollar Coin $53b Stablecoin Second largest stablecoin, pegged to

Name Market Cap.

(as of 5/27/22)

Type Description

(USDC) the U.S. dollar. Reserve assets held

in regulated U.S. banks.

Binance Coin

(BNB)

$50b Traditional Cryptocurrency issued by crypto

trading platform Binance.

Originally created as a token to pay

for discounted trades.

Ripple (XRP) $19b Traditional Commonly used for cross-

border transactions and

exchanges into various types of

fiat currencies.

Binance USD

(BUSD)

$18b Stablecoin Like BNB, BUSD is the stablecoin

version of Binance’s

cryptocurrency offering. Regulated

by the NYDFS.

Cardano (ADA) $16b Traditional Created by the founder of Ethereum,

Cardano uses smart contracts to

enable identity management.

Dogecoin (DOGE) $10b Traditional Created from a fork in the Bitcoin

blockchain, Dogecoin has

unlimited issuance. Can be used for

payments facilitation.

3.4.2 Nominal Dependent Variable

Volatility is the nominal dependent variable, and it was created by calculating the daily

volatility, or the statistical measure of the dispersion of return, for Bitcoin. Daily

volatility was calculated using the Garman-Klass (GK) volatility measure (Garman &

Klass, 1980) as shown in Equation 1.

𝑍 1 𝐻 2𝐶

�

Equation 1. Garman-Klass (GK) Volatility Measure

While several studies of cryptocurrency rely on Generalized AutoRegressive Conditional

= √ 𝑛 ∑ 2 [(𝑙𝑜𝑔 𝑖

�

𝑖

)−(2 log

(2)− 1) (log 𝑖

𝑂

𝑖

)]

Heteroskedasticity (GARCH) to measure volatility, recent studies suggest that GK is

more efficient in measuring cryptocurrency volatility (Brauneis & Mestel, 2018; Tan,

Chan, & Ng, 2020). To create a nominal variable for volatility, the top 20% were

classified as “high volatility” while the remaining 80% were classified as “low

volatility”.

Using KNIME Analytics Platform, the two data sources were combined via the

Joiner node. The join was configured as an inner join with both the top and bottom input

being the identifier derived from dates. No columns were excluded at the join stage. Next,

the raw and imputed spread values were excluded from the dataset as the newly imputed

dependent variable was derived from the spread values and including them in the models

would be redundant. The spread values, BTC_Spread_OpenClose and

BTC_Spread_OpenClose_flat, were removed using the Column Filter node configured

for manual selection and enforced exclusion. In addition, the daily high, low, and adjusted

close were excluded to focus the model on the spread values.

3.4.3 Predictor Variables (Studies 1a and 1b)

The following predictor variables pertain to Study 1a and Study 1b:

Commodities. This study evaluates cryptocurrency volatility compared to ten commodity

variables (Table 3). Most of the mainstream cryptocurrencies, including

Bitcoin and Ethereum, are considered to be commodities by the Commodity Futures

Trading Commission (CFTC) (Miller, 2022). However, studies show that since the

shortrun price of commodities like gold and oil are uncertain, cryptocurrencies like

Bitcoin do not have a supply-side uncertainty (Gronwald, 2019). For this reason, there is

academic debate as to whether cryptocurrencies are truly commodities.

Name Definition Freq. of

Data

Data Source

Daily

The first two variables, gold spot price and gold mining production, were

selected for their popular comparison to cryptocurrencies. While studies have examined

the relationship between gold prices and cryptocurrency prices and found little predictive

or behavioral relationships, gold is still viewed by many as the most common

comparative physical asset for cryptocurrencies. Gold is also an effective hedge against

the consumer price index (CPI), and thus seen as an effective hedge against inflation

(Bampinas & Panagiotidis, 2015).

The next eight variables, which include variables relating to crude oil prices and

production as well as natural gas prices and production, were selected due to their

relevance in the energy markets. Cryptocurrency mining, the process by which new coins

are minted and transactions are approved to the blockchain, is essential to the functioning

of a cryptocurrency. However, mining is energy intensive. Some estimate that Bitcoin

mining, for example, consumes as much energy as the city of Houston (Tabuchi, 2022).

Given the importance of crude oil and natural gas to energy prices, it was important to

include those commodity variables in the study.

Table 3. Commodities Considered in Studies 1a and 1b

Gold spot Spot price of gold in U.S. dollars. World Gold Council –

price Gold spot price

Crude oil spot Crude Oil in Dollars per Barrel, Daily U.S. Energy Information price Products in

Dollars per Gallon. Administration

(Cushing, OK)

Crude oil spot Crude oil in U.S. dollars per barrel, Daily U.S. Energy Information price (Europe

products in U.S. dollars per gallon. Administration

Brent)

Crude oil

production

Crude oil production in barrels, by

country.

Monthly U.S. Energy Information

Administration

Natural gas U.S. price in dollars of natural gas Monthly U.S. Energy Information

spot price

(residential)

delivered to residential consumers;

U.S. dollars per thousand cubic

feet.

Administration

Natural gas

spot price

(industrial)

U.S. price in dollars of natural gas

delivered to industrial consumers;

U.S. dollars per thousand cubic

feet.

Monthly U.S. Energy Information

Administration

Natural gas

spot price

(commercial

)

U.S. price in dollars of natural gas

delivered to commercial

consumers; U.S. dollars

per thousand cubic feet.

Monthly U.S. Energy Information

Administration

Natural gas

imports

U.S. natural gas imports in millions of

cubic feet.

Monthly U.S. Energy Information

Administration

Natural gas

exports

U.S. natural gas exports in millions

of cubic feet.

Monthly U.S. Energy Information

Administration

Securities. This study evaluates cryptocurrency prices compared to three

measures of securities (Table 4). These three indices of the U.S. and Americas stock

market were selected for their size and relevance in capturing fluctuations in the

securities market. Like commodities, these variables continue to be part of the ongoing

debate on cryptocurrency classification. The Securities and Exchange Commission

(SEC), for example, recently launched an investigation into cryptocurrency trading

platform Coinbase over questions that some cryptocurrencies listed in their platform

were unregistered securities (Beyoud & Versprille, 2022)

Table 4. Securities Considered in Studies 1a and 1b

Name Definition Freq. of

Data

Data Source

S&P 500 daily

closing price

Open, high, low, and close price of

the S&P 500 in U.S. dollars.

Daily WSJ Markets

Name Definition Freq. of

Data

Data Source

Daily

NASDAQ

Composite

Index daily

close price

Open, high, low, and close price of

the NASDAQ in U.S. dollars.

Daily WSJ Markets –

NASDAQ Composite

Index

Dow Jones

Industrial

Average daily

close price

Open, high, low, and close price of

the Dow Jones Industrial Average in

U.S. dollars.

Daily WSJ M arkets – Dow

Jones Industrial Average

Fiat currencies. This study evaluates cryptocurrency prices compared to fiat

currency variables (Table 5). These five fiat currencies were selected as they are ranked

as the five largest fiat currencies by market capitalization by CoinMarketCap.com. For

each fiat currency, there are two variables. The first is the daily open, high, low, and

close price for each currency. The second is the currencies M1, a variable that is

calculated differently for each currency but is intended to capture the total amount of the

currency in circulation. Note that for some currencies, the full 10-year data set is not

available.

Table 5. Fiat Currencies Considered in Studies 1a and 1b

U.S. Dollar Open, high, low, and close price of WSJ Markets – Dollar

(USD) the U.S. dollar index in U.S. Index

dollars.

Chinese Yuan Open, high, low, and close price of Daily WSJ Markets – Chinese

(CNY) the Chinese Yuan index in U.S. Yuan

dollars.

Yen (JPY) Open, high, low, and close price of

the Japanese Yen in U.S. dollars.

Daily WSJ Markets – Japanese

Yen

Euro (EUR) Open, high, low, and close price of

the Euro in U.S. dollars.

Daily WSJ Markets – Euro

British Pound

(GBP)

Open, high, low, and close price of

the British Pound in U.S. dollars.

Daily WSJ Market – British

Pound

M1 for U.S.

Dollar

(circulating

supply)

M1 consists of (1) currency outside

the U.S. Treasury, Federal Reserve

Banks, and the vaults of depository

institutions; (2) demand deposits at

commercial banks less cash items

in the process of collection and

Federal Reserve float; and (3) other

checkable deposits (OCDs).

Monthly Federal Reserve Bank of

St. Louis

M1 for

Chinese Yuan

(circulating

supply)

M1 comprises currency in

circulation plus demand deposits in

national currency of resident

nonbank non-government sectors

with the PBC and banking

institutions.

Monthly Federal Reserve Bank of

St. Louis - China

M1 for

Japanese Yen

(circulating

supply)

M1 comprises notes and coins in

circulation outside banking

corporations and demand and

savings deposits of households,

nonfinancial corporations, local

governments, securities

companies, Tanshi companies, and

some other financial corporations

such as securities finance

companies with banking

corporations in national currency.

Monthly Federal Reserve Bank of

St. Louis - Japan

M1 for Euro

(circulating

supply)

M1 comprises currency in

circulation and overnight deposits.

Monthly Federal Reserve Bank of

St. Louis – Euro Area

M1 for British

Pound

(circulating

M1 comprises currency in

circulation and overnight deposits.

Monthly Federal Reserve Bank of

St. Louis – United

Kington

supply)

Other economic indicators. This study evaluates cryptocurrency prices

compared to three other economic variables (Table 6). These variables are important

indicators of economic performance although they are not assets specifically. These

variables, Treasury bill rate, Consumer Price Index, and U.S. unemployment rate, are

included in the study to capture economic volatility that may not be reflected in the other

asset classes.

Table 6. Other Economic Indicators Considered in Studies 1a and 1b

Name Definition Freq. of

Data

Data Source

Treasury Note 4-week, 13-week, 26- week and

52-week coupon and bank discount

rate in U.S. dollars.

Daily U.S. Department of the

Treasury

Consumer

Price Index

(urban

consumers)

All items less food and energy in

U.S. city average (often used as a

proxy for inflation).

Monthly U.S. Bureau of Labor

Statistics

U.S.

unemployment

rate

Rate of unemployment

(percentage) for adults over 16

years of age.

Monthly U.S. Bureau of Labor

Statistics

3.4.4 Predictor Variables (Study 2)

The following set of predictor variables pertain to Study 2:

Blockchain Performance Variables. Study 2 looks intrinsically at the

performance of the underlying blockchains for Bitcoin. The 15 variables included for

analysis in this study pertain to the performance and efficiency of the Bitcoin blockchain

(Table 7). Studies show that consensus algorithms play an important role in the efficiency

of blockchain transactions (Mingxiao et al., 2017). Other studies suggest that financial

transactions, and thus, value is increased when blockchains perform more efficiently

(Cocco, Pinna, & Marchesi, 2017). This study assesses the effect of these variables on

Bitcoin price volatility to better understand the role of blockchain mechanics on

Bitcoin price volatility. These data are all publicly available and were downloaded

from blockchain.com.

Table 7. Blockchain Performance Variables Considered in Study 2

Name Definition Freq. of

Data

Average block size The average block size in MB. Daily

Hash rate The estimated number of tera hashes per second

(trillions of hashes per second).

Daily

Consensus difficulty A relative measure of how difficult it is to find a new

block. The difficulty is adjusted periodically as a

function of how much hashing power has been

deployed on the network.

Daily

Confirmed transactions

per day

The number of transactions processed each day. Daily

Transactions per block The number of transactions included in each block. Daily

Trade volume The total USD value of trading volume on major

bitcoin exchanges.

Daily

Mempool size The number of transactions waiting to be confirmed. Daily

Estimated transaction

volume

The total estimated value in USD of transactions on the

blockchain.

Daily

Total transaction fees

(USD)

The total USD value of all transaction fees paid to

miners.

Daily

Fees per transaction

(USD)

Average transaction fees in USD per transaction. Daily

Miners’ revenue (USD) Total value in USD of Coinbase block rewards and

transaction fees paid to miners.

Daily

Cost per transaction Miners’ revenue divided by the number of transactions. Daily

Median confirmation

time The median time for a transaction with miner fees to be

included in a mined block and added to the public ledger.

Daily

Market capitalization The total U.S. dollar market capitalization of a Daily

particular asset

Market price The average USD market price across major bitcoin Daily

exchanges.

3.4.5 Data Set Preparation

In accordance with the procedures outlined above in Section 3.3, data sets were

prepared for model development. For all three studies, a unique identifier derived from

record date is utilized as the primary key. For this study’s dependent variables (i.e.,

cryptocurrency price), the publicly available data includes prices for every single day of

the 10-years under evaluation. For some of the predictor variables, however, daily price

(or unit) data is not always available. For example, for the variables reflecting market data

(e.g., gold spot price) prices are available for every weekday, but not for Saturdays and

Sundays. In these cases, KNIME Analytics Platform estimated the values for the weekend

days using linear interpolation. For some of the predictors (e.g., natural gas imports),

relevant units are only available monthly. In those cases, the monthly value is used as a

proxy for every day within that month. In aggregate, Study 1a includes 3,805 records

(representing days) with 67 variables in each record for a total of 254,935 data points,

although some cryptocurrencies do not have the full 3,805 days of data due to their more

recent launch. Study 1b includes 3,805 records (representing days) with 64 variables in

each record for a total of 243,520 data points. Study 2 includes the same 3,805 records

(representing days) with 13 unique variables for a total of 49,465 data points.

As with Studies 1a and 1b, the numeric value for date is also utilized as the

primary key in Study 2. The input data is derived from two CSV files. The first is an

aggregation of the bitcoin blockchain performance data which includes average block

size, cost per transaction, estimated transaction volume, fees per transaction, hash rate,

median confirmation time, mempool size, miner revenue, network difficulty, number of

transactions, number of transactions per block, and trade volume. The second is an

aggregated set of data that includes daily open, high, low, close, adjusted close, volume,

and spread for each of the ten cryptocurrencies.

3.5 Testing and Evaluation

Various models were developed and tested that best suit the data and the research questions

posed in this study.

3.5.1 Prediction Models

Studies 1a and 1b look to see if cryptocurrency price volatility can be predicted

by blockchain performance variables. For this question, the following prediction models

were developed:

Decision Trees – This model evaluates the training data by recursively dividing

the data into several, cascading subsets based on a binary classification. Each classified

subset is assessed for its prediction accuracy. This division continues until the model is

dominated by one classification. The Gini Index is the most common method of assessing

the goodness of the split (Delen, 2021).

Ensembles Models – With this modeling approach, multiple analytics models are

combined into a single outcome to produce better predictions. According to Delen (2021,

p. 190), “rather than build models and then select the single best model, the ensemble

approach build many models (Figure 3) and uses them all for the task they are intended to

perform.”

Figure 3. Simple Taxonomy for Ensemble Models

•Random Forest – This model expands the prediction accuracy of

decision tree models by aggregating the results of several decision tree

models and creating a bootstrapping ensemble, also known as ‘bagging’.

The resulting bootstrapping ensemble is created by taking a random

subset of the input variables. As a result, random forest models do not use

all the available training data (on average, only 37% of training data is

used in bagging) (Abbott, 2014; Delen, 2021, p. 233). The results of

random forest models tend to outperform the results of standard decision

tree models since the prediction outcomes are determined from averages,

thus smoothing variations in the input data.

•Gradient Boosted Trees – This approach iteratively improves prediction

models by “boosting” records that were incorrectly classified in prior

iterations. In other words, the feedback from incorrectly predicted records

determines which records to boost in subsequent iterations to improve

prediction accuracy. Conceptually, Gradient Boosted Trees are a type of

neural network which is analogous to networks in the human brain that

can learn from past experiences. Neural networks learn from data without

rigid assumptions, and therefore, can be used to generalize (Delen, 2021).

•Multi-Layer Perceptron (MLP) – A multi-layer perceptron neural

network is known as “deep feedforward networks”. MLP networks can

process very large data sets in which multiple layers can be examined to

identify sophisticated patterns that might otherwise not be seen. In these

deep networks, no feedback connections are allowed.

•Naïve Bayes – This model employs a simple method of probability-based

classification. The model assumes that the input variables are not

independent, meaning that the presence or absence of the variables have

no impact on the presence or absence of the other variables (Delen, 2021).

Study 2 looked to see if Bitcoin price volatility can be predicted by blockchain

performance variables. For this question, the Naïve Bayes model was replaced with the

model below:

Logistic Regression – This is a data analysis technique that employes correlation

analysis used to identify relationships between two or more variables.

3.5.3 Prediction Accuracy

The performance of each model was assessed by the following metrics:

Accuracy – This reflects the ratio of True Positive (TP) and True Negatives (TN)

instances over the total number of instances which is represented by TP + TN + False

Positives (FP) + False Negatives (FN).

𝑇𝑃 + 𝑇𝑁

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =

𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁

Equation 2. Accuracy

Sensitivity (or True Positive Rate, TPR) – This is the ratio of TP over the total positives (which

includes TP and FN).

𝑇𝑃

𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =

𝑇𝑃 + 𝐹𝑁

Equation 3. Sensitivity

Specificity (or True Negative Rate, TNR) – This is the ratio of TN over the total

negatives (which includes TN and FP). The False Positive Rate (FPR) is calculated as 1 –

Specificity.

𝑇𝑁

𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =

𝑇𝑁 + 𝐹𝑃

Equation 4. Specificity

𝑇𝑁

𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑅𝑎𝑡𝑒 = 1 −

𝑇𝑁 + 𝐹𝑃

Equation 5. False Positive Rate

For all the models developed in this study, various methods were used to assess the

accuracy of the prediction findings (Delen, 2021):

Simple Split – This method randomly divides the data into at least two groups

including a training set and a validation set. The training data set is used to build the

algorithm that is later used to test the model’s prediction accuracy on the validation set.

k-Fold Cross-Validation – Given that the two (or more) data sets used in a Simple Spit

are rarely identical, k-Fold Cross-Validation attempts to minimize bias associated with

these differences. To do this, k-Fold Cross-Validation uses rotation estimation

methodology in which 1) the entire data set is divided into k subsets (or number of folds)

that are equal in size and mutually exclusive, 2) the model runs its training algorithm k

times, and 3) the prediction accuracy of the model is calculated by averaging the

individual prediction accuracies of the k models. k (or the number of folds) can be set to

any number, but the most used number is 10.

Area Under the Receiver Operating Characters (ROC) Curve (AUC) – This is

a graphical representation of prediction accuracy in which the TPR is plotted on the yaxis

and the FPR is plotted on the x-axis. This metric does reflect the prediction accuracy of an

individual model. Instead, it ranks how well each model performs. Models with higher

AUC values are likely to have higher accuracy values.

CHAPTER IV

IV. RESULTS

4.1 Introduction

To answer the research questions posed in the three studies, 50 models which

included the nominal dependent variable (i.e., volatility), the dependent variables (i.e.,

prices for nine cryptocurrencies), and the predictor variables were developed. In every

model, one or more of the predictor variables predicted volatility accuracy at a 90%

or greater accuracy rate by at least one type of machine learning model. Given that

90% accuracy represents a well-known standard for efficacious model prediction

(Delen, 2021), the results of these analyses are relevant and will inform our

understanding of cryptocurrency behavior.

The detailed models are presented in Appendix III; however, the sections below

present the ten summary models that illustrate the studies’ results and suggest findings

that will be explored in Chapter 5.

4.2 Study 1a – Predicting Cryptocurrency Price Volatility from Economic Variables

Study 1a asked the following questions:

Question 1a: How effective are advanced machine learning models in

predicting cryptocurrency price volatility?

Question 1b: Which variables are most significant in cryptocurrency

price volatility?

To answer these questions, the first models seek to understand whether the

economic predictor variables could accurately predict the volatility of cryptocurrency

prices. Figure 4 depicts the volatility of the six traditional cryptocurrencies since 2018.

(For reference, ADA = Cardano, BNB = Binance, BTC = Bitcoin, DOGE = Dogecoin,

ETH = Ethereum or Ether, and XRP = Ripple.) Given that Dogecoin (DOGE) clearly

represents outsized volatility in comparison to the other five cryptocurrencies, Figure 5

presents the same volatility trends but with DOGE removed to better visualize overall

volatility.

Greater volatility corresponds to specific time periods that appear consistent with

what is generally known about cryptocurrencies. For example, in the years prior to 2022,

the volatility of certain cryptocurrencies corresponds to the timeframe in which they were

launched. In 2022, the “crypto winter” impacted all cryptocurrencies, although Ripple

(XRP) and Binance (BNB) experienced the largest volatility.

Figure 4. GK Volatility of Cryptocurrencies (all)

600%

169.50%

%353.87

2315.98%

%952.15

%737.63

292.31%

500%

%1000

1500%

%2000

%2500

ADA BNB BTC DOGE ETH XRP

Figure 5. GK Volatility of Cryptocurrencies (without DOGE)

The next six models include all the economic predictor variables and the volatility

measure for each of the six traditional cryptocurrencies. Figure 6 below looks at all the

economic variables and their abilities to predict Cardano (ADA) volatility. As shown in

Table 8, the Random Forest model predicted Cardano (ADA) volatility with the highest

prediction accuracy (92.8%). Figure 7 ranks the importance of the predictor variables as

explained by the Random Forest model (the only model that ranks variable importance).

Interestingly, Crude Oil Production (a commodity) is the most important predictor of

Cardano’s (ADA) price volatility. The other ranked variables are some fiat currencies and

other cryptocurrency/stablecoin volumes and prices. The NASDAQ Daily Spread (a

security measure) appears to have some predictive ability, but it is low.

169.50%

353.87%

451.18%

292.31%

%100

200%

300%

400%

%500

ADA BNB BTC ETH XRP

Figure 6. ROC Curve for Cryptocurrency Volatility Prediction Models (ADA)

Table 8. Study 1a - ADA Volatility Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 90 359 26 25 0.898 0.783 0.932 0.068

Random Forest 97 367 18 18 0.928 0.843 0.953 0.047

Gradient

Boosted Trees 88 360 25 27 0.896 0.765 0.935 0.065

Multi-Layer

Perceptron

Neural Network

98 338 47 17 0.872 0.852 0.878 0.122

Naïve Bayes 90 252 133 25 0.684 0.783 0.655 0.345

Crude_Oil_Production _U.S.0.772

XRP_Adj_Close0.752

BTC_Volume

USDT_Volume

Euro_Prices_Close

ETH_Volume

XRP_Spread_OpenClose

SOL_Adj_Close

US_Dollar_Circulation

NASDAQ_Daily_Spread

Figure 7. Random Forest Variable Importance (ADA)

Figure 8 below looks at all the economic variables and their abilities to predict

Binance (BNB) volatility. As shown in Table 9, three models predicted Binance (BNB)

volatility with over 90% accuracy; however, the Random Forest model had the highest

prediction accuracy (96.5%). Table 9 ranks the importance of the predictor variables for

the Random Forest model. Again, a commodity, Natural Gas Production –Residential, is

the most important predictor. Other predictors of Binance’s (BNB) volatility include

some fiat currencies and other cryptocurrency/stablecoin volumes and prices. However,

another commodity, Natural Gas Imports, also shows predictive ability. As with Cardano

(XRP), a security measure, the Dow Jones Daily Spread, has some predictive ability, but

it is low (Figure 9).

0.564

0.526

0.41

0.385

0.371

0.31

0.299

0.287

Figure 8. ROC Curve for Cryptocurrency Volatility Prediction Models (BNB)

Table 9. Study 1a - BNB Volatility Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 87 366 17 12 0.940 0.879 0.956 0.044

Random Forest 94 371 12 5 0.965 0.949 0.969 0.031

Gradient

Boosted Trees 88 368 15 11 0.946 0.889 0.961 0.039

Multi-Layer

Perceptron

Neural Network

86 329 54 13 0.861 0.869 0.859 0.141

Naïve Bayes 53 283 100 46 0.697 0.535 0.739 0.261

Natural_Gas_Price_Residential0.85

US_Dollar_Circulation 0.798

XRP_Adj_Close

USDT_Volume

Natural_Gas_Imports

British_Pound_Value_Change

Euro_Prices_Close

Dow_Jones_Daily_Spread

DOGE_Spread_OpenClose

US_Dollar_Close

Figure 9. Random Forest Variable Importance (BNB)

Figure 10 below looks at all the economic variables and their abilities to predict

Bitcoin (BTC) volatility. As previously discussed, Bitcoin has the longest history

(launched in 2008) and is, by far, the largest cryptocurrency in terms of market

capitalization. As shown in Table 10, three models predicted Bitcoin’s (BTC) volatility

with over 90% accuracy; however, the Random Forest model had the highest prediction

accuracy (94.8%). Figure 11 ranks the importance of the predictor variables for the

Random Forest model. The top four predictors were all measures of fiat currency – with

the U.S. Dollar Circulation and U.S. Dollar Close as the top predictors. This finding raises

interesting questions about its potential classification relative to the other studied

cryptocurrencies which will be explored further in Chapter 5. A commodity, Natural Gas

Price –Residential, is also predictor of Bitcoin (BTC) volatility. The Dow Jones Daily

Spread, a security measure, has some predictive ability, but it is low.

0.66

0.517

0.505

0.49

0.413

0.378

0.356

0.306

Figure 10. ROC Curve for Cryptocurrency Volatility Prediction Models (BTC)

Table 10. Study 1a - BTC Volatility Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 81 359 23 19 0.913 0.810 0.940 0.060

Random Forest 91 366 16 9 0.948 0.910 0.958 0.042

Gradient

Boosted Trees 74 366 16 26 0.913 0.740 0.958 0.042

Multi-Layer

Perceptron

Neural Network

73 350 32 27 0.878 0.730 0.916 0.084

Naïve Bayes 53 293 89 47 0.718 0.530 0.767 0.233

Figure 11. Random Forest Variable Importance (BTC)

Figure 12 below looks at all the economic variables and their abilities to predict

Dogecoin (DOGE) volatility. As previously discussed, Dogecoin exhibited some of the

highest levels of volatility of any of the studied cryptocurrencies. As shown in Table 11,

the Random Forest model had the highest prediction accuracy (94.4%). Figure 13 ranks

the importance of the predictor variables for the Random Forest model. The top two

predictors were measures of other cryptocurrencies – Cardona (XRP) and Tether

(USDT). Two commodities, Natural Gas Price –Residential and Natural Gas Imports,

were also predictors of Dogecoin (DOGE) volatility. The remaining predictors were other

cryptocurrency and fiat currency measures.

0.31

0.365

0.456

0.47

0.48

0.484

0.543

0.635

Dow_Jones_Daily_Spread

DOGE_Adj_Close

USDC_Spread_OpenClose

BNB_Spread_OpenClose

Treasury_Rates_52_Weeks_Bank_Discount

Natural_Gas_Price_Residential

Euro_Prices_Close

Treasury_Rates_52_Weeks_Coupon

US_Dollar_Close

US_Dollar_Circulation

Figure 12. ROC Curve for Cryptocurrency Volatility Prediction Models (DOGE)

Table 11. Study 1a - DOGE Volatility Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 69 360 31 22 0.890 0.758 0.921 0.079

Random Forest 76 379 12 15 0.944 0.835 0.969 0.031

Gradient

Boosted Trees 68 364 27 23 0.896 0.747 0.931 0.069

Multi-Layer

Perceptron

Neural Network

65 343 48 26 0.846 0.714 0.877 0.123

Naïve Bayes 72 247 144 19 0.662 0.791 0.632 0.368

Japanese_Yen_Clos

Figure 13. Random Forest Variable Importance (DOGE)

Figure 14 below looks at all the economic variables and their abilities to

predict Ethereum (ETH) volatility. Launched in 2015, Ethereum is the second largest

cryptocurrency in terms of market capitalization. As shown in Table 12, two models

predicted Ethereum’s (ETH) volatility with over 90% accuracy; however, the Random

Forest model had the highest prediction accuracy (94.6%). Figure 15 ranks the

importance of the predictor variables for the Random Forest model. The top predictor

was a measure of fiat currency – U.S. Dollar Circulation. However, a commodity

measure, Natural Gas Price – Residential, exhibited nearly identical predictive

abilities. The remaining predictors of Ethereum (ETH) price volatility were measures

of commodities, securities, fiat currencies, and other cryptocurrencies.

0.351

0.393

0.439

0.485

0.51

0.535

0.605

0.765

0.799

0.807

BTC_Volume

US_Dollar_Close

Natural_Gas_Imports

Euro_Prices_Close

XRP_Volume

XRP_Spread_OpenClose

Natural_Gas_Price_Residential

USDT_Volume

XRP_Adj_Close

Figure 14. ROC Curve for Cryptocurrency Volatility Prediction Models (ETH)

Table 12. Study 1a - ETH Volatility Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 70 336 46 30 0.842 0.700 0.880 0.120

Random Forest 88 368 14 12 0.946 0.880 0.963 0.037

Gradient

Boosted Trees 83 369 13 17 0.938 0.830 0.966 0.034

Multi-Layer

Perceptron

Neural Network

78 302 80 22 0.788 0.780 0.791 0.209

Naïve Bayes 46 295 87 54 0.707 0.460 0.772 0.228

Figure 15. Random Forest Variable Importance (ETH)

Figure 16 below looks at all the economic predictor variables and their abilities to

predict Ripple (XRP) volatility. As shown in Table 13, the Random Forest model

predicted Ripple (XRP) volatility with the highest prediction accuracy (92.9%). Figure 17

ranks the importance of the predictor variables for the Random Forest model. Again,

Natural Gas Price – Residential, a commodity, is the most important predictor of Ripple’s

(XRP) price volatility. Interestingly, the other ranked variables are most often measures

of fiat currencies, although the stablecoin, Tether (USDT), also shows predictive

abilities.

0.297

0.303

0.336

0.362

0.365

0.43

0.471

0.854

0.917

0.92

SP500_Daily_Spread

USDT_Volume

Japanese_Yen_Daily_Spread

Euro_Prices_Close

USDC_Spread_OpenClose

Dow_Jones_Daily_Spread

Natural_Gas_Imports

XRP_Adj_Close

Natural_Gas_Price_Residential

US_Dollar_Circulation

Figure 16. ROC Curve for Cryptocurrency Volatility Prediction Models (XRP)

Table 13. Study 1a - XRP Volatility Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 83 347 31 21 0.892 0.798 0.918 0.082

Random Forest 90 358 20 14 0.929 0.865 0.947 0.053

Gradient

Boosted Trees 89 356 22 15 .923 0.856 0.942 0.058

Multi-Layer

Perceptron

Neural Network

87 340 38 17 0.886 0.837 0.899 0.101

Naïve Bayes 71 277 101 33 0.722 0.683 0.733 0.267

Figure 17. Random Forest Variable Importance (XRP)

4.2 Study 1b – Predicting Stablecoin Depegging from Economic Variables

Study 1b asked the following questions:

Question 2a: How effective are advanced machine learning models in

predicting the depegging of a stablecoin?

Question 2b: Which variables are most significant in predicting

stablecoin depegging?

Stablecoins are, by their very names, designed to mimic the performance of

specific fiat currencies (i.e., they are designed to be pegged to certain fiat currencies).

Study 1b seeks to understand which, if any, economic variables can predict depegging

behavior by considering the three largest stablecoins in terms of market capitalization.

(For reference, USDT = Tether, USDC = U.S. Dollar Coin, and BUSD = Binance USD.)

These three stablecoins were intended to be pegged to the U.S. dollar. For the purposes

of this study, depegging is defined as instances in which the price of the stablecoins fell

below $1.00 U.S.

0.351

0.389

0.404

0.454

0.46

0.477

0.628

0.65

0.806

1.009

Natural_Gas_Imports

Treasury_Rates_52_Weeks_Bank_Discount

Treasury_Rates_52_Weeks_Coupon

BTC_Volume

DOGE_Volume

Japanese_Yen_Close

Euro_Prices_Close

US_Dollar_Close

USDT_Volume

Natural_Gas_Price_Residential

Figure 18 below shows the depegging instances for Tether (USDT). As shown, the

price of Tether fell below the U.S. dollar numerous times, and particularly during the

2017 – 2021 period. As shown in Table 14, two models predicted Tether’s (USDT)

depegging with over 90% accuracy with the Random Forest model at a prediction

accuracy of 91.8% and the Gradient Boosted Trees at 91.9%. Figure 20 ranks the

importance of the predictor variables for the Random Forest model (the only model that

ranks variable importance). The top two predictors were cryptocurrencies – Solana

(SOL) and Ethereum (ETH). The NASDAQ close, a security measure, also showed

strong predictive abilities. The remaining predictors of Tether (USDT) depegging were

measures of fiat currencies and other cryptocurrencies.

Figure 18. Peg Value Scatterplot (USDT)

Figure 19. ROC Curve for Cryptocurrency Volatility Prediction Models (USDT)

Table 14. Study 1b - USDT Depegging Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 86 933 87 36 0.892 0.705 0.915 0.085

Random Forest 81 967 53 41 0.918 0.664 0.948 0.052

Gradient

Boosted Trees 82 968 52 40 0.919 0.672 0.949 0.051

Multi-Layer

Perceptron

Neural Network

102 844 176 20 0.828 0.836 0.827 0.173

Naïve Bayes 122 581 439 0 0.616 1 0.570 0.430

Figure 20. Random Forest Variable Importance (USDT)

Figure 21 below shows the depegging instances for U.S. Dollar Coin (USDC). As

shown, the price of Ethereum fell below the U.S. dollar numerous times, and particularly

during the initial years after its launch. As shown in Table 15, two models predicted

Ether’s (USDC) depegging with over 90% accuracy with the Random Forest model at a

prediction accuracy of 90.3% and the Gradient Boosted Trees model at 91.6%. Figure 23

ranks the importance of the predictor variables for the Random Forest model (the only

model that ranks variable importance). The top two predictors of U.S. Dollar Coin’s

(USDC) depegging were stablecoins – Binance USD (BUSD) and Tether (USDT). The

next two were measures related to the U.S. Dollar. Interestingly, three commodities

measures also showed predictive abilities.

0.403

0.433

0.475

0.49

0.494

0.563

0.614

0.735

0.987

1.003

XRP_Volume

Treasury_Rate_52_Weeks_Coupon

Crude_Oil_Production_U.S.

DOGE_Adj_Close

BTC_Volume

US_Dollar_M1

US_Dollar_Circulation

NASDAQ_Close

ETH_Adj_Close

SOL_Adj_Close

Figure 21. Peg Value Scatterplot (USDC)

Figure 22. ROC Curve for Cryptocurrency Volatility Prediction Models (USDC)

Table 15. Study 1b - USDC Depegging Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 73 932 88 49 0.880 0.598 0.914 0.086

Random Forest 98 933 87 24 0.903 0.803 0.915 0.085

Gradient

Boosted Trees 99 947 73 23 0.916 0.811 0.928 0.072

Multi-Layer

Perceptron

Neural Network

107 857 163 15 0.844 0.877 0.840 0.160

Naïve Bayes 115 775 245 7 0.779 0.943 0.760 0.240

Figure 23. Random Forest Variable Importance (USDC)

Figure 24 below shows the depegging instances for Binance USD (BUSD). As shown,

the price of Binance USD seemed to hover closer to the U.S. dollar than either

Tether or U.S. Dollar Coin. As shown in Table 16, three models predicted Binance USD

(BUSD) depegging with over 90% accuracy. The Decision Trees model had a prediction accuracy

of 91.8%, the Random Forest model had an accuracy of 92.6%, and the Gradient Boosted Trees

0.329

0.452

0.454

0.467

0.495

0.621

0.682

0.703

0.935

0.965

Treasury_Rates_8_Weeks_Coupon

BNB_Close

Treasury_Rates_8_Weeks_Bank_Discount

SOL_Volume

Natural_Gas_Exports

BTC_Volume

US_Dollar_Circulation

US_Dollar_M1

USDT_Volume

BUSD_Adj_Close

model was 93.2%. Figure 26 ranks the importance of the predictor variables for the Random

Forest model (the only model that ranks variable importance). The top four predictors of Binance

USD’s (BUSD) depegging were all measures related to the U.S. dollar. The remaining predictors

were securities (NASDAQ Close and SP500 Close), fiat currencies (Chinese Yuan), and other

cryptocurrencies. No commodities predicted Binance USD’s depegging.

Figure 24. Peg Value Scatterplot (BUSD)

Figure 25. ROC Curve for Cryptocurrency Volatility Prediction Models (BUSD)

Table 16. Study 1b - BUSD Depegging Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 92 956 55 39 0.918 0.702 0.946 0.054

Random Forest 106 951 60 25 0.926 0.809 0.941 0.059

Gradient

Boosted Trees 101 963 48 30 0.932 0.771 0.953 0.047

Multi-Layer

Perceptron

Neural Network

125 889 122 6 0.888 0.954 0.879 0.121

Naïve Bayes 131 856 155 0 0.864 1 0.847 0.153

US_Dollar_Circulation1.002 USDC_Adj_Close0.974

US_Dollar_M1

USDC_Volume

NASDAQ_Close

USDT_Volume

Chinese_Yuan_M1

SP500_Close

BTC_Volume

DOGE_Volume

Figure 26. Random Forest Variable Importance (BUSD)

4.3 Study 2 – Predicting Bitcoin Price Volatility from Blockchain Variables

Study 2 asked the following questions:

Question 3a: How effective are advanced machine learning models in

predicting price volatility of Bitcoin?

Question 3b: Which blockchain variables are most significant in

predicting Bitcoin price volatility?

Study 2 looked at price volatility in relation to the variables that underpin the

cryptocurrency’s blockchain as opposed to the economic variables evaluated in Studies

1a and 1b. Given that Bitcoin has the most longevity and thus, historical data, Study 2

analysis focused on Bitcoin price volatility. Figure 27 graphically depicts the price

volatility of Bitcoin. As shown in Table 17, two models predicted Bitcoin volatility with

over 90% accuracy. The Decision Trees model had a prediction accuracy of 90.6% and

0.883

0.754

0.64

0.529

0.523

0.484

0.386

0.351

the Gradient Boosted Trees model had an accuracy of 90.3%. Figure 29 ranks the

importance of the predictor variables for the Random Forest model (the only model that

ranks variable importance). The blockchain variable that exhibited the overwhelmingly

highest prediction accuracy was mempool size. Mempool size reflects the number of

transactions waiting to be confirmed to the blockchain. As such, it is a proxy for

resource availability (e.g., computing power and/or number of miners for whom

validation is profitable). This suggests that resource availability on the Bitcoin

blockchain is the largest single predictor of Bitcoin price volatility – an idea that will be

further explored in

Chapter 5.

Figure 27. GK Volatility of Bitcoin (2014 - 2022)

Figure 28. ROC Curve for Bitcoin Price Volatility Prediction Results

Table 17. Study 2 - Bitcoin Blockchain Model Prediction Results

Model TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Decision Trees 128 907 67 40 0.906 0.762 0.931 0.069

Random Forest 120 891 83 48 0.885 0.714 0.915 0.085

Gradient

Boosted Trees 130 901 73 38 0.903 0.774 0.925 0.075

Multi-Layer

Perceptron

Neural Network

115 876 98 53 0.868 0.685 0.899 0.101

Logistic

Regression 94 715 259 74 0.708 0.560 0.734 0.266

Figure 29. Random Forest Variable Importance for Bitcoin Volatility

0.003

0.134

0.159

0.255

0.321

0.344

0.35

0.388

0.416

0.444

0.447

1.058

Median_Confirmation_Time

Cost_Per_Transaction

n_Transactions

Estimated_Transaction_Volume

Miners_Revenue

Fees_Per_Transaction

n_Transactions_Per_Block

Network_Difficulty

Avg_Block_Size

Hash_Rate

Trade_Volume

Mempool_Size

CHAPTER V

V. DISCUSSION AND CONCLUSION

5.1 Introduction

The three studies presented in this paper resulted in models with high levels of

prediction accuracy. Study 1a looked at the predictive abilities of certain economic

variables on the price volatility of six traditional cryptocurrencies. Study 1b looked at the

predictive abilities of those same economic variables on the depegging of three

stablecoins (i.e., when the prices of those stablecoins fell below the U.S. dollar). Finally,

Study 2 looked at the predictive abilities of various blockchain variables on the price

volatility of Bitcoin.

All three studies were designed to better understand the volatility of

cryptocurrencies. Table 18 below summarizes the prediction results from each

cryptocurrency’s best volatility model. As shown, the Random Forest (RF) model had the

best accuracy results for all the traditional cryptocurrencies considered in Study 1a.

Gradient Boosted Trees (GBT) models had the best accuracy results for the stablecoins

considered in Study 1b and for the variables considered in Study 2.

Table 18. Top Volatility Model Prediction Results

Crypto-

currency

Model

Type

TP TN FP FN Accuracy Sensitivity Specificity

False

Positive

Rate

Cardano

(ADA) RF 97 367 18 18 .928 .843 .953 .047

Binance

(BNB) RF 94 371 12 5 .965 .949 .969 .031

Bitcoin (BTC) RF 91 366 16 9 .948 .910 .958 .042

Dogecoin

(DOGE) RF 76 379 12 15 .944 .835 .969 .031

Ethereum

(ETH) RF 88 368 14 12 .946 .880 .963 .037

Ripple (XRP) RF 90 358 20 14 .929 .865 .947 .053

Tether

(USDT) GBT 82 968 52 40 .919 .672 .949 .051

U.S. Dollar

Coin (USDC) GBT 99 947 73 23 .916 .811 .928 .072

Binance USD

(BUSD) GBT 101 963 48 30 .932 .771 .953 .047

Bitcoin

Blockchain GBT 130 901 73 38 .903 .774 .925 .075

Research questions 1a, 2a, and 3a specifically asked if advanced machine

learning models with the studied variables could accurately predict cryptocurrency price

volatility. Given that all three studies produced prediction models with high levels of

accuracy

(accuracy > 90%), these results positively answer those three research questions. More

importantly, they will inform both the academic and industry literature on this new field

of study.

5.2 Discussion

To best frame the findings in the three studies, it is helpful to begin with a discussion of the

results of Study 2.

5.2.1 Study 2 Discussion

The random forest prediction model results from Study 2 show that resource availability

or the cost of transactions is the most important variable for determining Bitcoin price

volatility. Figure 29 indicated that mempool size exhibits the best prediction ability on

Bitcoin price volatility. Given the mempool size is a proxy for resource availability (i.e.,

computing powers and/or number of miners for whom validation is profitable), the

results suggest that volatility in Bitcoin prices follows standard economic behaviors

associated with supply and demand.

Mempool size is a value reflecting the number of transactions waiting to be confirmed

and added to the blockchain. In proof-of-work blockchains, such as the Bitcoin

blockchain, transaction confirmation is executed by validator nodes (often called

“miners”). The results of Study 2 suggest that the most significant predictor of Bitcoin

volatility is the number of transactions that have been executed but not confirmed. This

implies that volatility increases when the margin on validation services is insufficient. On

the Bitcoin blockchain, the validation reward is fixed (the current reward is 6.25

Bitcoins), so it is reasonable to expect that the primary driver of mempool size is the cost

of validation. Since proof-of-work blockchains require validator nodes to calculate a

mathematical proof-of-work, validation is an energy intensive process. Therefore, the

cost of validation is likely correlated to the cost of energy.

5.2.2 Study 1a Discussion

As shown in Table 19 below, commodity variables were the top predictive

variable for most cryptocurrencies (4 of the 6). Interestingly, both Bitcoin and Ethereum

had fiat currencies as the top predictor variable (although the difference between the top

two predictor variables – one fiat currency and one commodity – in the Ethereum model

were negligible). Ethereum and Bitcoin are the two largest cryptocurrencies by market

capitalization, so the predictive power of fiat currencies in determining volatility is

worthy of further review. Except for DOGE, no other cryptocurrencies in the study had

another cryptocurrency (including Bitcoin) as the top predictive variable. Cryptocurrency

critics often cite the interconnectedness of cryptocurrencies as a driver of volatility. The

results from Study 1a do not support that conclusion.

While commodity variables appear to be the most important predictive variable

across the cryptocurrencies included in Study 1a, it is one specific type of commodity that

appears to drive importance – energy. Three of the four cryptocurrencies (BNB, XRP, and

ETH) where commodities were the top predictive variable had residential natural gas as

the top predictor variable. The fourth, ADA, had crude oil production as the top predictor

variable. As described in section 5.2.1 above, the results are in alignment with the results

from Study 2.

While not the direct focus of this study, it is worth noting that the natural gas

variable with the most predictive importance was residential consumption, not

commercial. These findings suggest that the miners themselves, or the entities

responsible for validating transactions on the blockchain, may be more diverse than

previously believed. It also suggests that there may be a significant volume of

mining activity occurring in residential locations.

Table 19. Top Predictive Variable (by class) for Traditional Cryptocurrencies

Traditional Other

currenciesCrypto- Stablecoins Securities Commodities CurrenciesFiat

EconomicIndicators

Cardano X

(ADA)

Binance

(BNB) X

Bitcoin

(BTC) X

Dogecoin X X

(DOGE)

Ethereum

(ETH) X X

Ripple (XRP) X

5.2.3 Study 1b Discussion

As shown in Table 20 below, depegging behavior of both Tether (USDT) and

U.S. Dollar Coin (USDC) was best predicted by other cryptocurrencies. The results from

Study 1b indicate a clear relationship between U.S. Dollar Coin and Binance USD as

both stablecoins appear high on the variable importance list for one another. Tether

depegging, a stablecoin with a long history of speculation that it is under-reserved, was

most predicted by Solana which may suggest a relationship between the two coins. U.S.

Dollar Coin, the newest and least volatile stablecoin, was the only stablecoin with a fiat

currency as the top predictor variable.

Table 20. Top Predictive Variable (by class) for Stablecoins

Other

TraditionalcurrenciesCrypto- Stablecoins Securities Commodities CurrenciesFiat

EconomicIndicators

Tether

(USDT) X

US Dollar

Coin X

(USDC)

Binance

USD X

(BUSD)

5.3 Conclusions

Since their inception, cryptocurrencies have been the source of substantial speculation.

Extreme price fluctuations, public collapses, and nefarious use cases all drive conjecture.

The three studies in this paper sought to illuminate the drivers of cryptocurrency behavior

by using machine learning techniques to evaluate the predictive properties of a wide

range of economic and blockchain performance variables. The results suggest that

volatility in cryptocurrency markets is subject to the traditional economic drivers of

supply and demand, however the supply and demand mechanisms themselves may be

unexpected and require further research.

Machine learning models do not lend themselves to discrete conclusions. Rather, the

models deployed in this study serve to predict volatility amongst the cryptocurrencies

and stablecoins assessed in the studies. The model results serve as observations that

imply relationships and directionality that should be assessed in further studies. It is also

important to remember that machine learning models rely on the collection of the

variables included within the model. Any conclusions regarding specific variables and

their importance within the models are for discussion purposes only as the full model

accuracy reflects the full ranges of variables. This section describes observations from

the models as well as inferences as to the root cause of certain model outcomes. Study

1a explored the economic drivers of volatility amongst the six largest traditional

cryptocurrencies. In general, it appears that commodities, not fiat currencies, securities,

or other cryptocurrencies, are the primary predictors of volatility for traditional

cryptocurrencies. These results are surprising given recent claims by the U. S. Securities

and Exchange Commission (SEC) that many cryptocurrencies, including Cardano

(ADA), Binance Coin (BNB), and Ripple (XRP) included in this study, are, in fact,

securities (U.S. Securities and Exchange Commission, 2023a, 2023b). However, given

the well-studied relationship between energy prices and cryptocurrency prices, it is

reasonable to expect that commodities, particularly oil and natural gas, would have

predictive properties for cryptocurrency volatility.

When assessed alongside the results of Study 2, the combined observation is that

commodities, and specifically oil and residential natural gas consumption, are the

primary predictors of mempool size and, thus, cryptocurrency volatility. While this

observation is not inherently new, it does provide enhanced clarity into the relationship

between energy, blockchain performance, and cryptocurrency volatility. These results

suggest that the triangulation between those three factors may be more significant than

previously believed. Considering these results, disruptions, or changes to energy

markets, such as the recent Russian invasion of Ukraine, may have an outsized impact on

blockchain performance and thus cryptocurrency prices. Similarly, significant changes to

blockchain consensus mechanisms, as in the case of Ethereum’s recent change to

proofof-stake consensus, may impact cryptocurrency volatility.

Perhaps the most interesting results come from Study 1b. Often marketed as a stable

alternative to highly volatile cryptocurrencies, stablecoins have positioned themselves

as the next generation of cryptocurrencies. Even sovereign states, including the United

States, are actively researching CBDCs, a form of state-backed stablecoin. Given this

attention, the observation that stablecoin depegging may not be predicted by the fiat

currencies that back them is novel. Oddly enough, it seems as though stablecoins are

more dependent on cryptocurrency markets than cryptocurrencies themselves. If or how

this may impact future CBDCs remains to be seen.

5.4 Limitations

While the three studies in this paper examined several predictor variables

(asset classes, economic indicators, and blockchain performance variables), the

number of potential predictor variables (even within the broad categories of variables

selected for this paper) are limitless. As such, the specific predictor variables selected

for these studies represent a natural inherent limitation. A discussion of some of the

potential variables that may be related to cryptocurrency price can be found in Chapter

II, Other Potential Influences on Cryptocurrencies.

The timing and potential impact of new or expanded regulatory oversite that

relates specifically to cryptocurrency activity is not captured in these studies. The

Securities and Exchanges Commission (SEC) and Commodity Futures Trading

Commission (CFTC) are still developing their positions on the asset classification of

cryptocurrencies. Further, the Office of the Comptroller of the Current (OCC) issued its

first interpretative letter for banks in 2020, while the Federal Deposit Insurance

Corporation (FDIC) and the Federal Reserve issued their letters in 2022. The downstream

impact of regulatory oversite on cryptocurrency behavior is a major opportunity for

future research, however another limitation in this paper is that the impact of new or

expanded regulatory oversite has not yet been realized.

As discussed in this paper, the two largest cryptocurrencies, Bitcoin and

Ethereum, utilize proof-of-work consensus protocols to validate blockchain transactions.

In September 2022, Ethereum moved to a proof-of-stake consensus protocol which is

different in many material respects from proof-of-work (Ethereum, 2022). The

downstream impact of Ethereum’s new blockchain validation approach is another

opportunity for future research as the predictive properties of the studied variables may

change with respect to Ethereum once the change from proof-of-work to proof-of-stake is

fully realized. Regardless, another limitation of this paper is that the studied variables

pre-date this change in Ethereum’s validation approach.

At the time of this paper’s publication, the volatility of cryptocurrency markets is

a well-observed phenomenon. A “crypto winter” is broadly defined as a period in which

most cryptocurrencies experience lower values (Chohan, 2022). While several crypto

winters have been observed in the past, the record prices of cryptocurrencies in 2020

made the crypto winter experienced in 2022 particularly obvious (and painful, for some)

(Chohan, 2022). As the scope of this paper includes data only to June 1, 2022, it is unclear

how the market downturn of the 2022 crypto winter may impact the efficacy of the

paper’s predictive models. Additionally, Study 2 relies on a calculation of daily volatility

for prediction. Predictive results may be biased based on the measure of volatility used in

the study.

Finally, certain global events have been observed to influence cryptocurrency

prices. For example, when facing economic sanctions, countries sometimes turn to crypto

markets to allow for cross-border payments. With the ongoing political events between

Ukraine and Russia, there is discussion that economic sanctions on Russia will encourage

more Russian activity in crypto markets, thus inflating the price of cryptocurrencies.

Another example was the sharp increase in cryptocurrency activity observed during the

Covid-19 pandemic, a time when U.S. interest rates were low and stock market

performance was volatile (Katsiampa et al., 2022). The analysis described in Chapter III

did not attempt to tease out the impact of these (or other) global events.

5.5 Recommendations for Future Research

These studies utilized data science techniques to explore

volatility in cryptocurrency markets. Future studies could expand upon these findings to

further evaluate the significance of the relationships overserved in these studies.

Additionally, future studies should explore the effectiveness of these models in predicting

the volatility from the 2022-2023 crypto winter as those data were not included in these

studies. From a methodological perspective, future models could expand upon the

models in these studies by incorporating additional variables that may further improve

the predictability and effectiveness of the models. For example, these studies did not

incorporate any natural language processing (NLP) techniques that may further explain

volatility in cryptocurrency markets. Future studies that combine the variables explored

in these studies along with NLP and other techniques may provide additional insights into

the behavior of cryptocurrency markets. Future studies that seek to study volatility may

also benefit from comparing several measures of volatility. Certain measures, like

GARCH, may provide additional perspective on cryptocurrency volatility.

From a variable inclusion and data standpoint, further studies could continue to explore

the details of additional cryptocurrency and commodity variables. Future studies should

research the significance of commodities in driving cryptocurrency prices. Such studies

may assist legislators and regulators in determining the appropriate classification for

cryptocurrencies and other similar assets.

Students also viewed