Surveying the prediction of risks in cryptocurrency investments using recurrent neural networks

Rihab Qasim Abdulkadhim; Hasanen S. Abdullah; Mustafa Jasim Hadi

doi:10.1515/eng-2022-0509

Artikel Open Access

Surveying the prediction of risks in cryptocurrency investments using recurrent neural networks

Rihab Qasim Abdulkadhim , Hasanen S. Abdullah und Mustafa Jasim Hadi

Veröffentlicht/Copyright: 29. Januar 2024

Veröffentlicht von

Veröffentlichen auch Sie bei De Gruyter Brill

Manuskript einreichen Informationen für Autor*innen

Aus der Zeitschrift Open Engineering Band 14 Heft 1

Abstract

Decentralized cryptocurrencies have received much attention over the last few years. Bitcoin (BTC) has enabled straight online expenditures without the need for centralized financial institutions. Cryptocurrencies are used not only for online payments but are also increasingly used as financial assets. With the rise in the number of cryptocurrencies, including BTC, Ethereum (ETH), and Ripple (XRP), and the millions of daily trades through different exchange services, cryptocurrency trading is prone to challenges similar to those seen in the traditional financial industry, such as price and trend forecasting, volatility forecasting, portfolio building, and fraud detection. This study examines the use of Recurrent neural networks (RNNs) for predicting BTC, ETH, and XRP prices. Accurate price prediction is essential for investors and traders in this volatile market. Machine learning techniques, including RNNs, Long-Short-Term Memory (LSTM), and convolutional neural networks, have been employed to forecast cryptocurrency prices with varying degrees of success. The aim of this study is to evaluate the effectiveness of RNNs in predicting cryptocurrency prices and compare their performance with other established methods. The results indicate that RNNs, particularly LSTMs and Gated Recurrent Units, demonstrate excellent capabilities in accurately predicting currency prices and providing insights to investors and traders in the cryptocurrency market.

Keywords: cryptocurrencies; RNN; LSTM; risk of cryptocurrencies; bitcoin

1 Introduction

Satoshi Nakamoto created Bitcoin (BTC) in 2009, using a pseudonym. A decentralized digital currency system called BTC exists [1]. Virtual currency is a comparatively recent phenomenon in terms of world finance. Its identity, construction, and purpose are continually changing, but it is becoming more and more acknowledged as a superior medium for global finance with enormous potential [2]. BTC has been developed to eliminate the need for time-consuming and expensive money transfer fees, as well as reliance on trustworthy third parties like banks, credit cards, and governments [3]. Due to its enormous potential, BTC distinguishes itself from other virtual currencies. Clones of BTC or other cryptocurrencies make up the majority of cryptocurrencies [4].

Blockchain technology is a decentralized and secure method of recording digital transactions that eliminates the need for a trusted third party to validate and safeguard transactions. BTC, the first and most well-known application of blockchain technology, uses a cryptographic proof system to execute transactions between two parties. Every transaction in the BTC network is broadcast to every node before being verified and then recorded in a public ledger [5]. The value of a cryptocurrency can be affected by the number of users, with more users leading to higher prices [6]. While they offer benefits such as faster transactions, lower fees, and increased privacy, they also come with risks such as price volatility, cyber-attacks, and a lack of regulation [7]. Another concern is the use of cryptocurrencies for illicit activities [8]. Since then, over 4,000 alternate cryptocurrencies such as Ethereum (ETH) and Ripple (XRP) have been created, proving that the cryptocurrency market has emerged in the financial area. BTC , ETH, and XRP are the most widely used cryptocurrencies, as they hold almost 79.5% of the global cryptocurrency market capitalization [3]. Deep learning (DL) refers to powerful Machine learning (ML) algorithms that specialize in solving nonlinear and complex problems by exploiting large amounts of data to become efficient predictor models, and it has proven to be a reliable tool for analysis [9,10].

The study of data gathering, analysis, interpretation, forecasting, visualization, and organization is known as statistics. ML is utilized efficiently in numerous fields, including forecasting [11], real-time advertising, malware detection [12], and text-based sentiment analysis [13]. Recurrent neural networks (RNNs) have emerged as highly effective models in the field of deep learning, demonstrating remarkable performance in recent years. These models are specifically designed to tackle sequential data by leveraging a memory-based architecture [14].

In this research, we will be studying recent literature on the prediction of currency prices, which is considered a major risk factor in currency trading. The studies will be presented in detail in Section 2. We will learn about the most important risks and features of currencies, take a quick look at RNNs, make comparisons between studies, analyze the results, and discuss them.

2 Literature review

This study presents a comparative analysis of published studies on predicting BTC prices using neural networks (NNs) and statistical methods. The selected publications were chosen from recent years, and the analysis was conducted by considering the frequency of the dataset, the causal association, and the methodology used in each study

Dutta et al. have been using a combination of exogenous and endogenous variables to forecast the price of BTC. The authors proposed two DL models, RNN and Gated recurrent units (GRU), to make BTC price predictions. The study found that GRU was more effective in learning and training BTC price changes, and both RNN and Long-Short-Term Memory (LSTM) were slower in comparison. The root mean square error (RMSE) values for the training dataset were NN 0.02, LSTM 0.010, and GRU 0.010, and for the test dataset, they were NN 0.31, LSTM 0.024, and GRU 0.019. The study concluded that DL models like LSTM and GRU outperform traditional ML models [15].
Jay et al. introduced layer-wise randomness into the activations of observed features in multilayer perceptron (MLP) and LSTM to simulate market volatility. Their best-case trial, which used 23 features and a window size of 7 days, resulted in an average improvement of no more than 5% in Mean Absolute Percentage Error (MAPE) [16].
Othman et al. used LSTM to create a model for predicting the BTC stock market. The study concludes by discussing future work based on the methodology used and techniques and tools employed to predict BTC’s price on Yahoo Finance’s stock market. According to their prediction, BTC’s price will exceed USD 12,600 with an RMSE of 288.59866 in the days following the prediction [17].
Cavalli and Amoretti have introduced a new method for predicting BTC trends using the 1D convolutional neural network (CNN). Their method includes combining social media data, blockchain transaction history, and financial metrics to construct datasets, and using a cloud-based system to collect and analyze large amounts of data. The 1D CNN model they created achieved a prediction accuracy of 74.2%, which is reported to be better than LSTM models. They also presented a trading strategy that maximizes profit during bullish trends and minimizes losses during bearish trends based on their model. Their contribution to the field of BTC trend prediction is significant [18].
Yang suggested a method that combines Ensemble Empirical Mode Decomposition (EEMD) and LSTM to investigate the subject of BTC price forecasting for the following day. Using the EEMD-LSTM approach requires additional work, as indicated by the outcome [19].
Sebastião and Godinho explored the predictability of BTC, ETH, and Litecoin using ML approaches. They validated and tested their models during a period of high volatility and bear markets to evaluate the accuracy of their predictions. They utilized trading and network activity data from August 15, 2015, through March 3, 2019, with the test sample starting on April 13, 2018. Five out of 18 different models had success rates of less than 50% during the test period. However, the ensemble of five models that generate identical signals performed well for ETH and Litecoin, with high positive ratios and annual rate returns. The study concludes that ML can be used to develop profitable trading strategies in cryptocurrency markets, even during tough times in the market [20].
Schulte and Eggert aim to find markers for BTC price fluctuations and create an hourly price prediction model using a trained LSTM. They identify major price prediction signs and compare the LSTM results with other techniques, demonstrating an impressive accuracy rate of 96.48%. The study utilized 47 input features over 10 months, leading to discussions on the use of NNs in predicting stock prices. The research makes valuable contributions to the field of BTC price prediction and suggests the potential of LSTM models for accurate and reliable predictions [21].
Jaquart et al. investigate the predictability of the BTC market across different time horizons using ML models. Technical aspects are found to be the most important for most techniques, followed by blockchain-based and sentiment/interest-based elements. Predictability improves with longer horizons, and a long-short trading strategy based on quantiles can yield up to 39% monthly profits before transaction costs, but the accuracy of predictions is around 50%. RNNs and gradient-boosting classifiers perform well in the predictive task, outperforming a random classifier [22].
Das et al. proposed a method for predicting BTC prices using a combination of DL techniques, involving CNN, LSTM, Bi-LSTM, RNN, and Bi-RNN. While individual models like CNN, LSTM, Bi-LSTM, RNN, Bi-RNN, or autoregressive integrated moving average achieve satisfactory results, the hybrid method proposed in the study outperforms them with RMSE, MAPE, and mean absolute error (MAE) values of 2.69, 1.78, 2.20, and 1.29, respectively. The study demonstrates the efficacy of combining different DL building blocks for more reliable BTC price prediction [23].
Aljojo et al. used a nonlinear autoregressive exogenous NN model to investigate the impact of timestamps on BTC’s value. They find that transaction timestamps significantly influence BTC’s performance variance with 96% forecast accuracy. Certain transaction events show repeating patterns, allowing for the anticipation of BTC’s price with less accuracy. Incorrect timestamps can cause uncertainty and have a negative impact on the BTC market. The study highlights the importance of accurate time stamping for participants in the BTC market [24].
Andi’s study used LSTM ML to predict BTC prices using a normalized dataset. The study found that using a combination of ML techniques leads to better results. The accuracy of the prediction model was found to be 97.2% [25].
Buslim et al.’s study contrasts the effectiveness of GRU, RNN, and LSTM models using a dataset from August 17, 2017 to April 13, 2021. Grid Search and Random Search are utilized to determine the best-performing model. The results show that GRU with Grid Search has the best performance, with MAE training of 0.0043 and a testing of 0.0594 [26].
Zhang et al. combined SDAB, which uses the advanced deep neural network model (SDAE), with an effective integration method to predict price integration. The study used various exogenous variables and found that SDAE-B was more accurate and had fewer errors compared to traditional ML methods such as Least Squares Support Vector Machine (LSSVM) and back propagation. The SDAE-B method had a MAPE of 0.016, RMSE of 131.643, and DA of 0.817, making it a more effective method in handling the randomness and nonlinearity of the BTC price [27].
Zhang et al. proposed an LSTM prediction (LSTM-P) neural network model for BTC and gold price prediction. To smooth out price data fluctuations, a noise decrease technique based on the wavelet transform has been used, and an optimized LSTM-P model has been created utilizing historical gold and BTC price data. With a MAPE of 6.08% for LSTM-P and 4.81% for standard LSTM, the LSTM-P model outperforms standard LSTM models and other time series forecasting models [28].
Luo et al. used ML and multiscale analysis to create ensemble prediction models by matching various techniques to their respective multiscale components. BTC price data from 2017/11/24 to 2020/4/21 and 2020/4/22 to 2020/11/27 were used for training and prediction. The ensemble models achieved 95.12% forecast accuracy and outperformed benchmark models. The suggested models were robust in both rising and falling market conditions, and various algorithms were applied to components with different time scales [29].
Kang et al. suggested a hybrid DL model, 1D CNN-GRU, for forecasting cryptocurrency prices. The model combines a 1D CNN and stacked GRU to encode BTC price data over time and capture its long-range dependencies. The execution of the 1D CNN-GRU model has been evaluated on BTC, ETH, and XRP datasets, and it was found to outperform previous approaches with RMSE values of 43.933 for BTC, 3.511 for ETH, and 0.00128 for XRP [30].
Aljadani’s study applies Bidirectional LSTM and GRU-based DL algorithms to three Yahoo Finance public real-time BTC datasets. GRU outperforms Bi-LSTM for BTC, ETH, and Cardano, with the GRU model achieving the lowest RMSE values of 0.01711, 0.02662, and 0.00852 for the three cryptocurrencies, respectively. The proposed framework achieves significant performance in predicting cryptocurrency prices [31].
Nematallah et al. proposed using DL processes such as RNNs and LSTMs to predict BTC market price trends. The models are evaluated based on their MAPE and RMSE values for predicting the next 15, 30, and 60 days. The LSTM model is found to be the superior method with a lower MAPE of 3.01% compared to RNN’s MAPE of 7.10%, even though it takes longer to assemble [32].
Mehta and Sasikala’s research focusses on using ML techniques to forecast BTC prices. By analyzing daily data for 1,681 cryptocurrencies between November 2015 and April 2018, the authors show that their approach, which employs simple trading techniques supported by cutting-edge ML algorithms, outperforms traditional benchmarks. The study’s goal is to forecast the cryptocurrency market’s short-term future and provide real-time BTC price predictions. The use of LSTM RNN on the available dataset yielded an impressive 92% accuracy [33].

3 Characteristics of cryptocurrencies

Cryptocurrencies offer many benefits, such as decentralization, security, and privacy, but they also have some downsides, such as high volatility and susceptibility to hacking and fraud. Users can evaluate whether these currencies are suitable for their financial objectives by assessing each feature listed in Table 1.

Table 1

Important features of cryptocurrencies

Feature	Details	Source
Decentralization	Cryptocurrencies are characterized by decentralization where there is no intermediary or entity that controls payment and transfer operations	Böhme et al. [41]
Security	Cryptocurrencies provide high security for users, as payment transactions are protected using advanced encryption techniques	Antonopoulos [42]
Privacy	Cryptocurrencies allow complete privacy for users, as no personal information is recorded when making payment transactions	Böhme et al. [41]
Accessibility	Cryptocurrencies enable easy access to payment and transfer services	Antonopoulos [42]
Enhancing international transfers	Cryptocurrencies help enhance international transfer operations due to their ease of use	Antonopoulos [42]
Improving transparency	Cryptocurrencies provide great transparency in payment and transfer operations, as all transactions are recorded in an encrypted and transparent ledger	Narayanan et al. [43]
Hedge against inflation	Cryptocurrencies are a good option for hedging against traditional currency inflation	Böhme et al. [41]
High volatility	Cryptocurrencies are characterized by high volatility in their value, as their value can be significantly affected by changes in the global market, which can result in financial losses for users	Narayanan et al. [43]
High probability of hacking	Despite the security offered by cryptocurrencies, there is still a high probability of hacking due to vulnerabilities in software or user errors	Narayanan et al. [43]

Table 1 introduces the characteristics of cryptocurrencies, which are digital assets that support online payments and transfers. Due to the encryption of transactions, cryptocurrencies are characterized by decentralization and high security. They provide users with privacy and easy access to financial services. They can also improve international transfers and act as a hedge against inflation. However, their high volatility and vulnerability to hacking calls for precautionary measures. Overall, cryptocurrencies represent a significant development in the field of digital finance.

4 Risk of cryptocurrency

Cryptocurrencies, as an emerging asset class, present unique risks for investors and traders. Understanding these risks is essential for making informed decisions and mitigating potential losses. The following points highlight the most significant risks associated with investing in cryptocurrencies and provide insights for navigating the volatile digital currency landscape.

Investing in cryptocurrencies carries several risks, one of which is hype risk. This occurs when investors make impulsive decisions without fully understanding the technology or market dynamics, leading to significant losses when the market self-corrects. Therefore, thorough market analysis is crucial before investing, and the risks involved must be considered.
Another significant risk is a security risk, which stems from the high risk of scams, hacking, and theft in the cryptocurrency market. Investors must choose secure blockchain protocols, exchanges, and wallets to safeguard their investments.
Cryptocurrencies are known for their high volatility, which poses a considerable risk for investors. Having a long-term investing plan and being equipped to deal with volatility are both crucial.
Liquidity risk is also a concern, particularly for lesser-known altcoins, which may not have enough buyers or sellers, making it challenging to sell investments quickly at a fair price.
Vanishing risk refers to the possibility of cryptocurrencies failing and becoming worthless, resulting in significant losses for investors. Therefore, investors should conduct thorough research and invest in promising and established projects.
Regulatory risk is also a possibility as regulatory changes can impact the market and create volatility as cryptocurrencies become more mainstream. Keeping up with legislative changes is essential so that investing plans may be modified.
Finally, tax risk includes the possibility of cryptocurrency investments being subject to capital gains tax, and tax authorities becoming more stringent in enforcing tax regulations. Complying with tax regulations and seeking professional advice can help avoid potential penalties and losses [34].

5 Artificial neural network (ANN)

There are some similarities between ANNs and biological brains. Data are received, processed, and transmitted from and to nearby neurons by a group of simple signal processing units (artificial neurons). In the training phase, when ANN learns from the dataset given, the strength (weighting) of the connections among neurons is determined. ANNs are a vast category of algorithms. The main classes for supervised learning include [35] (Figures 1–3):

MLP network – This has layers of artificial neurons that are normally fully coupled. Data are transmitted among the levels in a feed-forward fashion. ANNs’ inputs are vectors. These ANNs are primarily employed for situations involving organized or unstructured data of limited complexity [36].
CNN – CNN is one of the DL techniques. A network that organizes artificial neurons in feed-forward layers that are often thinly linked. Convolution layers enable the use of multi-dimensional data (tensor matrices) and the detection of complicated patterns. As a result, such ANNs are mainly well adapted to huge issues involving unstructured data [37,38].
RNNs are a form of ANN capable of handling sequential data such as text, speech, and time series data. They use memory cells to store information from the past, allowing them to carry information forward through time. The best type of RNN for price prediction depends on the specific task and data quality [39].

Figure 1

MLP neural network structure [36].

Figure 2

Convolutional neural networks [38].

Figure 3

The simple RNN design schematic [39].

There are three types of RNNs:

Elman RNN (ERNN) is a form of RNN architecture that is often considered the most basic version of RNNs. ERNNs can be used for various applications such as natural language processing, short-term load forecasting, tourist arrival forecasting, and electric load time series prediction. ERNNs have also been used in hybrid algorithms that combine particle swarm optimization and evolutional calculation to overcome the local minimum issues of gradient-based methods. The architecture of an ERNN consists of input layers, hidden layers, and output layers. Input and output layers have feedforward connections, while hidden layers contain recurrent connections. At each step, the input layer processes the component of a serial input, which can be real values, discrete values, one-hot vectors, and so on. The internal state of the network from the former time period is also processed in the hidden layer, which applies a nonlinear transformation using an activation function such as a sigmoid or hyperbolic tangent. The output of the network at each time step is calculated through a linear transformation on the matrix of the output weights applied to the current state of the network [40].
LSTM has a more advanced architecture than ERNN, which was first suggested by Hochreiter and Schmid Huber in 1997. LSTM is commonly used for accurately modeling short- and long-term dependencies in data. In LSTM, the vanishing gradient issue is resolved by avoiding recent observation bias and maintaining constant error flow backward in time. Five different nonlinear components that interact with one another in a specific way make up a cell, the more intricate internal processing unit that LSTM implements. Each component is controlled by three gates, each of which is realized by a pointwise multiplication and a sigmoid. For the purpose of completing a target task, the gates are trained using gradient descent. As well as image tagging, handwriting recognition, speech recognition, music composition, and grammar learning, LSTM has been used in many other series learning applications [40].
GRU is a different gated architecture that captures dependencies in an adaptive manner at various time scales. The amount that each hidden unit can remember or forget is adaptively controlled by the GRU by combining the forget and input gates into a single update gate. Due to the absence of an output gate-like control mechanism in the GRU, the internal state is always fully exposed to output. GRUs have been tested on various tasks such as statistical machine translation, anticipating the digits of the sum or difference of two input numbers, forecasting the next character in a synthetic XML dataset, and predicting polyphonic music. When using naive initialization, the results presented showed that GRU outperformed LSTM on most tasks with the exclusion of language modeling [40].

6 Comparative study of cryptocurrency forecasting methods

We made a comparison between the research on the basis of the method used, the type of data, and the results obtained, and this comparison includes three tables. Table 2 shows the application of several ML methods for predicting cryptocurrency prices, particularly BTC. The primary methods used are LSTM networks, RNNs, and CNNs. Diverse datasets were used, taken from various sources such as Kaggle, historical data, and other publicly available sources.

Table 2

Comparative analysis of cryptocurrency price prediction models based on accuracy

Authors	Year	Method	Dataset	Accuracy (%)
Mehta and Sasikala [33]	2020	LSTM, RNN	BTC prediction process	92
Cavalli and Amoretti [18]	2020	CNN	Different datasets	74.2
Schulte and Eggert [21]	2021	LSTM	Collected	96.48
Jaquart et al. [22]	2021	LSTM	Collected	>50
Aljojo et al. [24]	2021	Neural net	Kaggle BTC historical data	96
Andi [25]	2021	LSTM	Acquired dataset BTC price	97.2
Luo et al. [29]	2022	LSTM-ELM	Collected	95.12

It appears that the LSTM method is the most common among researchers for predicting BTC prices, with several studies achieving high accuracy rates using this method. On the other hand, a lower accuracy of 74.2% was achieved in the study that used CNN compared to LSTM studies. This difference in accuracy might indicate that LSTM could be more suitable for predicting BTC prices because of its ability to capture long-term dependencies in the data. There remains a need for further research and experimentation to determine the optimal method for predicting cryptocurrency prices and improve the overall accuracy of these predictions.

Table 3 compares the Mean Absolute Percentage Error (MAPE) for different methods of predicting the price of cryptocurrencies. The studies included primarily focused on LSTM and RNN methods applied to different datasets, with data collected from various sources or taken from Kaggle. The studies presented are from 2022, showing the most recent approaches to predicting cryptocurrency prices.

Table 3

Comparison of MAPE for various cryptocurrency price prediction methods

Authors	Year	Method	Dataset	MAPE
Zhang et al. [27]	2022	LSTM	Collected	6.08%
Fekry et al. [32]	2022	RNN, LSTM	Kaggle 2012 to 31 March 2021	7.10%, 3.01%

Table 3, suggests that LSTM methods have demonstrated better MAPE values for predicting cryptocurrency prices compared to other methods, particularly in the study by Fekry et al. However, it is important to consider that factors such as dataset choice and prediction window duration can significantly affect model performance. Further research should explore alternative ML techniques and examine the influence of various factors on prediction accuracy.

Table 4 compares various ML techniques used for cryptocurrency price prediction, concentrating on RMSE values for BTC, ETH, and XRP. The methods employed across the studies include LSTM, RNN, NN, CNN, Bi-LSTM, Bi-RNN, 1D CNN-GRU, GRU, and Bi-LSTM. The datasets used for these predictions were collected from various sources, such as Yahoo Finance and cryptocurrency exchange websites.

Table 4

Comparison of RMSE for cryptocurrency price prediction using different models

Authors	Year	Method	Dataset	RMSE (BTH)	RMSE (ETH)	RMSE (XRP)
Ferdiansyah et al. [17]	2019	LSTM, RNN	Collected	288.59866	—	—
Dutta et al. [15]	2020	NN, LSTM	Collected	0.031	0.024	—
Das et al. [23]	2021	CNN, LSTM, Bi-LSTM, RNN, Bi-RNN	Collected from Yahoo Finance	2.69%	1.78%	2.20%
Kang et al. [30]	2022	1D CNN-GRU	Collected from cryptocurrency exchange website	43.933	3.511	0.00128
Buslim et al. [26]	2021	GRU	Datasets from 17 August 2017 to 13 April 2021	0.0043	—	—
Aljadani [31]	2022	Bi-LSTM, GRU	Yahoo Finance	0.01711	0.02662	0.00852

From Table 4, it can be observed that there is no universally best method for all three cryptocurrencies, as different techniques yield varying RMSE values. For instance, Kang et al.’s 1D CNN-GRU model produced the lowest RMSE for ETH and XRP, while Dutta et al.’s NN and LSTM models achieved the lowest RMSE for BTC. These findings highlight the importance of selecting the appropriate ML method based on the specific cryptocurrency being predicted. Additionally, the results may be influenced by the chosen datasets and other factors, warranting further exploration to determine the most effective techniques for each cryptocurrency.

7 Discussion

The three tables presented provide a comparative analysis of various ML methods employed to predict cryptocurrency prices, particularly BTC, ETH, and XRP. According to Table 2, LSTM is the method most commonly employed and achieves more accurate results than CNN. Table 3 shows that LSTM methods have better MAPE values, and Table 4 demonstrates that different ML methods produce varying RMSE values for various cryptocurrencies when used to predict their prices. It is, therefore, crucial to consider factors such as the choice of dataset and duration of the prediction window underscoring the importance of selecting an appropriate method based on the specific cryptocurrencies being predicted. Overall, these findings suggest further research to determine the most effective machine-learning techniques for predicting cryptocurrency prices.

8 Conclusion

In this study, a thorough review of the numerous studies using NN approaches for cryptocurrency is given. The use of ML techniques, particularly RNNs, has shown promise in predicting the prices of cryptocurrencies such as BTC, ETH, and XRP. Researchers have identified that LSTM and GRU-based RNNs can achieve high accuracy rates and lower MAPE values, highlighting their potential for providing valuable insights for investors and traders in the rapidly evolving digital currency market. However, all factors affecting the cryptocurrency market, such as social networking site sentiments, and political and economic events that may have negative or positive impacts, must be studied. Further research is needed to determine the most effective ML techniques for predicting cryptocurrency prices and to fully comprehend the risks and benefits of this emerging asset category.

In our future work, we propose that one approach with great potential is to develop hybrid models that combine LSTM, DNN, and GRU architectures that promise to improve risk prediction in digital currency investments. These models leverage the strengths of each architecture to improve accuracy in the dynamic crypto market. LSTMs excel at capturing long-term dependencies, DNNs handle complex models, and GRUs provide the right balance between storage capacity and processing power. By combining these capabilities, hybrid models can extract valuable insights from different datasets, allowing investors to assess risk accurately and make informed decisions. Comparative analysis of the LSTM-DNN and GRU-DNN models helps determine the optimal architecture. Hybrid models represent a significant step forward in understanding market dynamics, mitigating risk, and making informed investment decisions. The integration of LSTM, DNN, and GRU into hybrid models offers a promising direction for research for predicting risk and empowering investors to act in the changing crypto market.

Conflict of interest: The authors state no conflict of interest.
Data availability statement: Most datasets generated and analyzed in this study are in this submitted manuscript. The other datasets are available on reasonable request from the corresponding author with the attached information.

References

[1] Rahouti M, Xiong K, Ghani N. Bitcoin concepts, threats, and machine-learning security solutions. IEEE Access. 2018;6:67189–205.10.1109/ACCESS.2018.2874539Suche in Google Scholar

[2] Abboushi S. Global virtual currency – Brief overview. J Appl Bus Econ. 2017;19(6).Suche in Google Scholar

[3] Olvera-Juarez D, Huerta-Manzanilla E. Forecasting bitcoin pricing with hybrid models: A review of the literature. Int J Adv Eng Res Sci. 2019;6(9):161–4.10.22161/ijaers.69.18Suche in Google Scholar

[4] Rauchs M, Hileman G. Global cryptocurrency benchmarking study. Cambridge Centre for Alternative Finance Reports; 2017.10.2139/ssrn.2965436Suche in Google Scholar

[5] Crosby M, Pattanayak P, Verma S, Kalyanaraman V. Blockchain technology: Beyond bitcoin. Appl Innov. 2016;2(6–10):71.Suche in Google Scholar

[6] Gandal N, Hamrick JT, Moore T, Oberman T. Price manipulation in the Bitcoin ecosystem. J Monetary Econ. 2018;95:86–96.10.1016/j.jmoneco.2017.12.004Suche in Google Scholar

[7] Bouri E, Molnár P, Azzi G, Roubaud D, Hagfors LI. On the hedge and safe haven properties of Bitcoin: Is it really more than a diversifier? Financ Res Lett. 2017;20:192–8.10.1016/j.frl.2016.09.025Suche in Google Scholar

[8] Yermack D. Is Bitcoin a real currency? An economic appraisal. In Handbook of digital currency. Academic Press; 2015. p. 31–43.10.1016/B978-0-12-802117-0.00002-3Suche in Google Scholar

[9] Ji S, Kim J, Im H. A comparative study of bitcoin price prediction using deep learning. Mathematics. 2019;7(10):898.10.3390/math7100898Suche in Google Scholar

[10] Kortus T, Krüger T, Gühring G, Lente K. Automated robust interpretation of intraoperative electrophysiological signals – A Bayesian deep learning approach. Curr Directions Biomed Eng. 2021;7(2):69–72.10.1515/cdbme-2021-2018Suche in Google Scholar

[11] Lahmiri S, Bekiros S. Cryptocurrency forecasting with deep learning chaotic neural networks. Chaos, Solitons Fractals. 2019;118:35–40.10.1016/j.chaos.2018.11.014Suche in Google Scholar

[12] Narudin FA, Feizollah A, Anuar NB, Gani A. Evaluation of machine learning classifiers for mobile malware detection. Soft Comput. 2016;20:343–57.10.1007/s00500-014-1511-6Suche in Google Scholar

[13] Long F, Zhou K, Ou W. Sentiment analysis of text based on bidirectional LSTM with multi-head attention. IEEE Access. 2019;7:141960–9.10.1109/ACCESS.2019.2942614Suche in Google Scholar

[14] Kostadinov S. Recurrent Neural Networks with Python Quick Start Guide: Sequential learning and language modeling with TensorFlow. Packt Publishing Ltd; 2018.Suche in Google Scholar

[15] Dutta A, Kumar S, Basu M. A gated recurrent unit approach to Bitcoin price prediction. J Risk Financial Manag. 2020;13(2):23.10.3390/jrfm13020023Suche in Google Scholar

[16] Jay P, Kalariya V, Parmar P, Tanwar S, Kumar N, Alazab M. Stochastic neural networks for cryptocurrency price prediction. IEEE Access. 2020;8:82804–18.10.1109/ACCESS.2020.2990659Suche in Google Scholar

[17] Ferdiansyah F, Othman SH, Radzi RZRM, Stiawan D, Sazaki Y, Ependi U. A LSTM-method for bitcoin price prediction: A case study yahoo finance stock market. In 2019 International Conference on Electrical Engineering and Computer Science (ICECOS). IEEE; 2019, October. p. 206–10.10.1109/ICECOS47637.2019.8984499Suche in Google Scholar

[18] Cavalli S, Amoretti M. CNN-based multivariate data analysis for bitcoin trend prediction. Appl Soft Comput. 2021;101:107065.10.1016/j.asoc.2020.107065Suche in Google Scholar

[19] Yang L. Next-day bitcoin price forecast based on artificial intelligence methods. arXiv preprint arXiv:2106.12961. 2021 Jun 21.Suche in Google Scholar

[20] Sebastião H, Godinho P. Forecasting and trading cryptocurrencies with machine learning under changing market conditions. Financial Innov. 2021;7(1):1–30.10.1186/s40854-020-00217-xSuche in Google Scholar PubMed PubMed Central

[21] Schulte M, Eggert M. Predicting hourly bitcoin prices based on long short-term memory neural networks. In Innov Inf Sys-tems: Volume II: A Collect Latest Res Technol Issues. Switzerland: Springer International Publishing; 2021. p. 754–69.10.1007/978-3-030-86797-3_50Suche in Google Scholar

[22] Jaquart P, Dann D, Weinhardt C. Short-term bitcoin market prediction via machine learning. J Financ Data Sci. 2021;7:45–66.10.1016/j.jfds.2021.03.001Suche in Google Scholar

[23] Das S, Billah M, Mumu SA. A hybrid approach for predicting Bitcoin price using Bi-LSTM and Bi-RNN based neural network. Lect Notes Comput Sci (including Subser Lect Notes Artif Intell Lect Notes Bioinforma). 2021;13113 LNCS:223–33. 10.1007/978-3-030-91608-4_23.Suche in Google Scholar

[24] Aljojo N, Alshutayri A, Aldhahri E, Almandeel S, Zainol A. A nonlinear autoregressive exogenous (NARX) neural network model for the prediction of timestamp influence on bitcoin value. IEEE Access. 2021;9:148611–24.10.1109/ACCESS.2021.3124629Suche in Google Scholar

[25] Andi HK. An accurate Bitcoin price prediction using logistic regression with LSTM machine learning model. J Soft Comput Paradig. 2021;3(3):205–17.10.36548/jscp.2021.3.006Suche in Google Scholar

[26] Buslim N, Rahmatullah IL, Setyawan BA, Alamsyah A. Comparing Bitcoin’s prediction model using GRU, RNN, and LSTM by hyperparameter optimization grid search and random search. In 2021 9th Inter-national Conference on Cyber and IT Service Management (CITSM). IEEE; 2021, September. p. 1–6.10.1109/CITSM52892.2021.9588947Suche in Google Scholar

[27] Zhang S, Li M, Yan C. The empirical analysis of Bitcoin price prediction based on deep learning integration method. Comput Intell Neurosci. 2022;2022:1–9.10.1155/2022/1265837Suche in Google Scholar PubMed PubMed Central

[28] Zhang X, Zhang L, Zhou Q, Jin X. A novel Bitcoin and gold prices prediction method using an LSTM-P neural network model. Comput Intell Neurosci. 2022;2022:1–12.10.1155/2022/1643413Suche in Google Scholar PubMed PubMed Central

[29] Luo C, Pan L, Chen B, Xu H. Bitcoin price forecasting: An integrated approach using hybrid LSTM-ELM models. Math Probl Eng. 2022;2022:1–7.10.1155/2022/2126518Suche in Google Scholar

[30] Kang CY, Lee CP, Lim KM. Cryptocurrency price prediction with convolutional neural network and stacked gated recurrent unit. Data. 2022;7(11):149.10.3390/data7110149Suche in Google Scholar

[31] Aljadani A. DLCP2F: a DL-based cryptocurrency price prediction framework. Discov Artif Intell. 2022;2(1):20.10.1007/s44163-022-00036-2Suche in Google Scholar

[32] Nematallah HF, Sedky AAH, Mahar KM. Bitcoin price trend prediction using deep neural network. Webology (ISSN: 1735-188X). 2022;19(3):77746–64.Suche in Google Scholar

[33] Mehta P, Sasikala DE. Prediction of Bitcoin using recurrent neural network. Int J Recent Technol Eng (IJRTE). 2020;8(6):1303–7.10.35940/ijrte.F7808.038620Suche in Google Scholar

[34] Danial K, Laurence T, Kent P, Bain T, Solomon MG. Cryptocurrency All-in-one for Dummies. United Kingdom: John Wiley & Sons; 2021.Suche in Google Scholar

[35] Guillod T, Papamanolis P, Kolar JW. Artificial neural network (ANN) based fast and accurate inductor modeling and design. IEEE Open J Power Electron. 2020;1:284–99.10.1109/OJPEL.2020.3012777Suche in Google Scholar

[36] Heidari AA, Faris H, Mirjalili S, Aljarah I, Mafarja M. Ant lion optimizer: theory, literature review, and application in multi-layer perceptron neural networks. Nature-Inspired Optimizers: Theories, Lit Rev Appl. 2020;811:23–46.10.1007/978-3-030-12127-3_3Suche in Google Scholar

[37] Cengiz E, Kelek MM, Oğuz Y, Yılmaz C. Classification of breast cancer with deep learning from noisy images using wavelet transform. Biomed Eng/Biomed Technik. 2022;67(2):143–50.10.1515/bmt-2021-0163Suche in Google Scholar PubMed

[38] Lu J, Tan L, Jiang H. Review on convolutional neural network (CNN) applied to plant leaf disease classification. Agriculture. 2021;11(8):707.10.3390/agriculture11080707Suche in Google Scholar

[39] Sherstinsky A. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. Phys D: Nonlinear Phenom. 2020;404:132306.10.1016/j.physd.2019.132306Suche in Google Scholar

[40] Bianchi FM, Maiorino E, Kampffmeyer MC, Rizzi A, Jenssen R. Recurrent neural networks for short-term load forecasting: an overview and comparative analysis. Cham, Switzerland: Springer; 2017.10.1007/978-3-319-70338-1Suche in Google Scholar

[41] Böhme R, Christin N, Edelman B, Moore T. Bitcoin: Economics, technology, and governance. J Econ Perspect. 2015;29(2):213–38.10.1257/jep.29.2.213Suche in Google Scholar

[42] Antonopoulos AM. Mastering bitcoin: unlocking digital cryptocurrencies. “O’Reilly Media, Inc.”; 2014.Suche in Google Scholar

[43] Narayanan A, Bonneau J, Felten E, Miller A, Goldfeder S. Bitcoin and cryptocurrency technologies: a comprehensive introduction. United States: Princeton University Press; 2016.Suche in Google Scholar

Received: 2023-05-18

Revised: 2023-07-24

Accepted: 2023-08-17

Published Online: 2024-01-29

This work is licensed under the Creative Commons Attribution 4.0 International License.

Artikel in diesem Heft

https://doi.org/10.1515/eng-2022-0509

Schlagwörter für diesen Artikel

cryptocurrencies; RNN; LSTM; risk of cryptocurrencies; bitcoin

Creative Commons

BY 4.0