Survey of Cryptocurrency Volatility Prediction Literature Using Artificial Neural Networks

We start by presenting a short description of the concept of cryptocurrency and the history behind it. Recently-developed literature that attempt to predict volatilities of cryptocurrency valuations through creation of hybrid artificial neural network models are then discussed. For the major part of the paper, we delve into details of multiple hybrid artificial neural networks that were thoroughly implemented to predict cryptocurrency volatilities. Results are reported within the form of a survey. Finally, we compare different methods and discuss their results follow at the end.


Introduction
The last decade has witnessed an enormous hike in usage and popularity of cryptocurrencies. The initial idea of a virtual currency was given by American computer scientist and cryptographer David Chaum in early 80's. His initiative first being creation of a secure signature system for credit card authorization, his research led to the idea of virtual wallets and currencies. In late 80's he founded the first virtual currency, ecash, which went bankrupt in a decade. ISSN 2162-4860 2022 The idea of a decentralized cryptocurrency emerged by Nakomoto in the original 2008 Bitcoin paper. Records indicate that the first Bitcoin transactions were performed in early 2009, despite the Nakomoto's character/pseudonym having remained unknown to date. Since 2009, All cryptocurrencies, including Bitcoin, have become more and more popular over time. The primary enormous price hikes of Bitcoin, as a result of a sudden increase in demand with supply not increasing equivalently, took place in 2017. Since then, all cryptocurrencies, especially Bitcoin, have been experiencing sudden hikes repetitively.

Business and Economic Research
Such price hikes make Bitcoin and most other cryptocurrencies extremely volatile financial assets in a holistic time-series view. Indeed, short-term investors are enormously interested in forecasting the timelines of such volatilities in order to maximize investment profits. Economic and Financial theory argues that volatilities are due to real-world shocks and/or supply and demand difference surges. However, recent time-series Bitcoin prices show extreme volatilities that may not be explained by classical financial and economic theory.
In recent time-series econometrics literature, there exist research papers that successfully attempt to model and forecast such unexplained volatility (e.g., Andersen et. al., 2003). More recently, cryptocurrency-specific literature has attempted to use artificial neural network methods to predict such volatility. Due to the novel nature of cryptocurrency data prediction research, the literature is thin and has just recently (after 2017) began growing. Throughout this paper, we explain the various methodologies and results used in artificial neural network models predicting cryptocurrency volatility. In conclusion, we compare results and elaborate on available paths for further research. It should be noted that this survey paper doesn't analyze literature focusing on predicting cryptocurrency prices. For a survey of predicting cryptocurrency prices, refer to Charandabi and Kamyar, 2021A.

Predictability of Volatility
Realized volatility is the time aggregation (normally over a trading day) of the sum of the squares of logarithmic returns on a highly frequent sample. Andersen et. al. (2003) devises a model to predict financial market volatilities in a specific framework. The model is established on the quadratic variance of the theory of arbitrage-free price processes in time series, and it uses connections among realized volatility and matrix of conditional covariance. The model is further tested on German and Japanese currency to USD exchange rates (on an autoregressive Gaussian vector for logarithmic values) and reportedly performs plausibly. The statistical particular details of the model benchmark involve a parametric mixture of lognormal and normal distribution. The mathematical baseline for Realized Volatility model devised by Andersen et. al. are given in relations (1) through (3), with a N-times sampled aggregation period between i and i + T periods.

Using Long Short Term Memory (LSTM)
A number of projects have attempted to predict Bitcoin price volatilities in varieties of ANN-LSTM settings using different data sets. Formulas written below depict the theoretical framework for any node. (Zhengyang et al., 2019) in which Architectural overview of mechanism of Long Short Term Memory is depicted in figure 2. ISSN 2162-4860 2022

Other Artificial Neural Network Frameworks
Among other literature using ANN methods to predict movements is Pabuç cu et al. (2020). Methods used include Support Vector Machine (SVM), Naï ve Bayes (NB), Random Forest (RF), and Logistic Regression (LR). The Support Vector Machine method is a non-stochastic method based on a process of minimizing structural risk through maximization of the margin between samples (negative, positive). The Naï ve Bayes method is a simplistic method based on the Bayes conditional probability principle Theorem. Allowing to compare results with other algorithms, The Random Forest method is an ensemble-learning algorithm based on a single classifier's incapability of determining the class of test data. Finally, the Logistic Regression is a popular technique modeling the probability of discrete outcomes. For each of the mentioned algorithms, multiple three-parameter combinations were determined and the most plausible of each was chosen for the final test. Table 2 shows the comparison of these methods, with True Positive, False Positive, ROC, and F statistic. The results are further ranked in terms of efficiency, using LR as the benchmark. Also establishing on Bayesian Probabilistic theory and principle, Jang et al. (2017) attempts to explain the volatility by proposing an artificial neural network with a multi-layer perceptron maximizing the value of posterior. They use a Bayesian Neural Networks (BNN) algorithm, taking blockchain data spanning from 2011 through 2017 into account, controlling for simultaneous values of stocks prices and currency exchange rates. Their efficiency results are given in Table 3. Given the resulting MAPE value of 1%, their attempt appears to be successful. ISSN 2162-4860 2022 Jang et al. (2017).

Autoregressive Model with Jumps
The autoregressive model with jumps was first introduced in Andersen et al. (2007). It uses the same realized volatility and logarithmic return as in relations (1) and (2), and accounts for bipower variation measures and corresponding nonparametric tests for jumps. Pichl et al. (2017) applies the Anderson methodology to the case of forecasting Bitcoin volatility. The regression equation is given in relation (6), with J values as jumps defined in Andersen et al. (2007). Table 4 shows the regression results for exchange rate of Bitcoin to USD, using 88 days of volatile price data in 2017 with daily 5-min time series. Based on the results, the authors conclude that their proposed regression model captures the dynamics of daily Realized Volatility as aggregated on the 5-minute grid plausibly. √ +1 = + 1 √ + 2 √ -5 + 3 √ -1 + 4 √ + 4 √ + 5 √ -5 + 6 √ -1 (6)

Discussion and Conclusion
As shown in the previous sections, the purpose of this survey paper is to present and compare the existing literature employing multiple artificial neural network-based approaches to predict cryptocurrency price volatilities. Noting the pros and cons of each of the methods presented (in terms of time elapsed, prediction accuracy, RMSE, MAPE, and R-squared), it appears that in a holistic manner the methods used in Jang et al. (2017), and the Ridge Regression method in Miuri et al. (2019) are of significant efficiency.
Due to the novel nature of this literature, every paper written in the area spans solely to the last three years. Therefore, we may expect many additional papers be written in the years to come. There exist some aspects of cryptocurrency price volatility forecasting with no current existent research papers. An area of possible further work, for example, is to focus solely on implementing data from less volatile cryptocurrencies (e.g. Ethereum, XRP, Chainlink, etc.) ISSN 2162-4860 2022 into the same frameworks. Indeed, many other frameworks could be implemented to forecast and/or measure the prediction efficiencies as well. For instance, the methodology in Filabadi and Mahmoudzadeh (2019) can be implemented to yield a novel model.
Because of the eminent recent hike in the popularity and price of cryptocurrencies (including but not limited to Bitcoin), introduction to the background, history, and technicalities of cryptocurrency is of critical importance to policymakers and other economic agents. The first transaction using bitcoins was to buy a pizza pie a decade ago, which would have been worth over $5 million today. With such quick development, all agents need to engage as fast as possible (Charandabi and Kamyar, 2021A).