CHAPTER TWO

LITERATURE REVIEW

Review of Related Work

For predicting the price of cryptocurrencies, there are numerous comparative studies available. Many
researchers have conducted cryptocurrency price predictions using a variety of methods. Few of them
employed DL models, while some just took ML techniques into account when conducting their research.
Researchers have also taken into account various cryptocurrencies, like Ethereum and Litecoin. Few
articles considered a short-term strategy, while others focused on long-term analysis. A few of these
papers are discussed below:

Phaladisailoed et al. Used 1-minute interval trade data from January 1, 2012, to January 8, 2018, via the
Kaggle website to predict Bitcoin values using the LSTM, GRU, Theil-Sen, and Huber regression models.
Theil-Sen and Huber and other regression models were tested using the scikit-learn and Keras libraries,
and the findings revealed that GRU and LSTM performed superior to them. With an of 0.00002 and an R2
of 0.992, GRU had the best accuracy, however, it took more time than Huber regression. They have solely
used R2 metrics to assess their models. Evaluating using only linear regression models makes them unable
to compute non-linear relations between variables.

Seabe et al. used three different types of RNN, including LSTM, GRU, and Bi-Directional LSTM. The study
focuses on Bitcoin, ethereum, and Litecoin, three of the most popular cryptocurrencies, and measures the
performance of the models using Mean Absolute Percentage Error (MAPE) and RMSE. According to the
experimental findings, Bi-LSTM fared better than LSTM and GRU in terms of prediction accuracy, with the
lowest MAPE values for Bitcoin, Litecoin, and Ethereum being 0.036, 0.041, and 0.124, respectively.
However, they did not take into account any ensemble machine learning models and instead just
compared deep learning algorithms. Only RMSE and MAPE measures were used to evaluate their models.

Vanderbilt et al. Tested 3 distinct cryptocurrencies — Bitcoin, Litecoin, and Ripple—using 3 different RNN
models—Simple RNN, LSTM, and GRU. From January 2015 to March 2020, they used historical price
information from coinmarketcap.com. In an effort to improve the model’s accuracy, they have also run
further tests using new data from Google Trends for the same time period. They came to the conclusion
that the three procedures were comparable in accuracy to one another. Additionally, the accuracy did not
improve further after conducting additional experiments utilizing the data from Google Trends. They
simply used the RMSE statistic to evaluate their model, and they subsequently tried to improve the output
by training them with each cryptocurrency’s daily closing price including or excluding Google Trends
information. To improve the validity of the forecast results, other evaluation metrics along with RMSE
were used in this thesis.

Kim et al. have tested on three cryptocurrencies namely Bitcoin, Ethereum, and Litecoin using LSTM and
GRU deep learning algorithms. The datasets were chosen from coinmarket based on kurtosis and
skewness. LSTM and GRU models are trained and tested on the same hyperparameter configuration while
increasing the number of epochs from 1 to 30. The accuracy of each model is measured by RMSE and MAE.
After the testing phase, they concluded that GRU was more advantageous for the downward stabilization
trend, and the LSTM was suitable for the upward stabilization trend.

Iqbal et al. focused on finding the most efficient technique out of ARIMA, FB Prophet, and XGBoosting for
predicting the future price of Bitcoin based on RMSE, MAE, and R2 parameters. After their
experimentation on all three algorithmic techniques, they found out that the parametric score of ARIMA is
the best of all three considered techniques. They have used the Kaggle dataset’s historical price data and
pre-processed that data and used it to build their models for comparison. They have mostly focused on
supervised machine learning algorithms and have not considered any other type of algorithms for their
comparison. On the other hand, our comparative analysis is based on a mixture of both Supervised
learning models RF, GB and RNN models LSTM, GRU.

In order to determine the future trends of the Ethereum cryptocurrency, Kumar et al. used deep learning
techniques including the Multi-Layer Perceptron (MLP) and LSTM, and evaluated their findings using the
RMSE, MAE, and metrics. They have made a long-term prediction and computed values for each day, hour,
and minute using previous pricing data. They used daily data that included open price, closing price, high
and low, volume per day, and hourly data from August 2015 to August 2018. They came to the conclusion
that the LSTM model is superior to the MLP in terms of robustness and accuracy for long-term reliance.
Our models were trained using the dataset that contained the technical indicators EMA, SMA, RSI based
on the historical price data.

Derbentsev et al. [16] focused on the short-term prediction model for price prediction of cryptocurrencies.
The updated Binary Auto Regressive Tree (BART) was adapted to series data and standard models. Study
shows that BART is more accurate than ARIMA. ARIMA model in slow-growing and transitional dynamic
times. RMSE for this algorithm for the horizon of 14, 21, and 30 days was within the ranges 4%, 6%, and 8%
respectively. They have taken closing prices from January 2017 to March 2019 according to yahoo finance,
and calculated their time series in log return. They have made the log return for the next time period their
target variable. The forecast was carried out on five different time horizons: 5, 10, 14, 21, and 30 days
using three models for each cryptocurrency. In contrast, we have considered the values SMA, EMA, and
RSI values as our target variables.

Chen et al. have identified Bitcoin prices through regular prices and high frequency prices to predict
Bitcoin prices through ML techniques at different frequencies. In comparison with the usual price
benchmark results, XGB and SDA machine learning algorithms have higher results with the highest
statistical accuracy of 66% and 65.3% respectively. They have employed 2 datasets: The first includes the
aggregated Bitcoin daily price, with a big interval and small scale, from CoinMarketCap.com. The second
dataset consists of 5-minute interval Bitcoin real-time trading price data at high-frequency and large scale
pulled from Binance. They have considered 12 features like block size, hash rate, etc, and used data from 2
different datasets to conduct their study whereas we have used technical indicators like (SMA, EMA, and
RSI) and a single dataset from Coinbase to conduct our experiment on.

Summary of related work results are noted in Table

Author
Technique Outcome Evaluation Dataset
RMSE:
ARIMA, Bitcoin price ARIMA: 322.4
Phaladisailo FBProphet, FBProphet:229.5 Kaggle
ed
prediction
XGB XGB: 369
 R2:
XBG: 4.855%
Returns:
Forecasting OLS:
returns for 3703 Equal-weighted:
cryptocurrencies, 6.297%
Various
Seabe OLS, XGB creating Capital-weighted:
sources
equal-weighted 1.852%
portfolio XGB:
Equal-weighted:
7.105%
Capital-weighted:
2.165%

Ensemble learning
methods, K-NN Close price of Accuracy:
model, gradient cryptocurrency and Ensemble Coin-
Vanderbilt boosted trees, their learning: 92.4% marketcap,
neural net model. corresponding Gradient Cci30.com
index boosted: 90%

Average:
Bitcoin price
Kim Decision trees Geometric: 2.77 Investing
prediction
Arithmetic: 4.20
Cryptocurrency Accuracy: Blockchain,
Iqbal LSTM, Regres- price prediction: Bitcoin: 0.9957
sion trees Etherscan
Bitcoin,Ethereum Ethereum: 0.9999
Accuracy:
LR: 66.0%
LSTM, QDA, LDA: 63.9% Coin-
Kumar SVM, RF, Bitcoin price QDA: 55.1% marketcap
XGB, LR, LDA. prediction SVM: 65.3% Binance
RF: 51.0%
XGB: 48.3%
LSTM: 57.0%
Cryptocurrency RMSE of BART:
ARIMA, forecasting: 14 days: 4% Yahoo
Derbentsev Regression tree, Bitcoin,Ethereum, 21 days: 6% finance
BART Ripple 30 days: 8%