πŸ“Introduction and related works

Bitcoin, the first decentralized cryptocurrency developed by Satoshi Nakamoto, has become the most valuable digital currency in the world. With the significant trading volume of cryptocurrencies, numerous types of digital currencies have emerged in the crypto world. Some well-known cryptocurrencies include Ethereum and Ripple, alongside other emerging coins such as Binance Coin, Solana, Cardano, and so on. This research focuses on cryptocurrency price prediction and profitable trading signal identification. Cryptocurrency price prediction is a time series problem that can be addressed through the use of deep learning techniques. While predicting cryptocurrency prices presents a challenge, developing price prediction algorithms is worthwhile as it plays a crucial role for cryptocurrency traders. The predicted prices are compared against actual prices, and the root mean square error is calculated. The main contributions of this paper are discussed in relation to the relevant research in the field:

  • Historical price data of cryptocurrencies are obtained from cryptocurrency exchanges. This research leverages data at different time intervals to improve price predictions (1 day, 12 hours, 4 hours, 2 hours, and 1 hour).

  • Feature expansion is performed on the historical price data of cryptocurrencies through normalization. Additionally, the data is preprocessed to remove missing values that may affect model learning. The cleaned data is then divided into training and testing sets for model learning and price prediction.

  • Deep learning architectures, particularly CNN, have been quite successful in stock market prediction due to their ability to extract high-level features more efficiently than simple artificial neural networks (ANN). Di Persio L, Honchar O (2016) used CNN with raw price data as a single feature while disregarding other technical indicators and achieved an accuracy of 53.6% in classification. Their study lacked the incorporation of multiple features in their ensemble models. Gunduz H, Yaslan Y, Cataltepe Z (2017) built upon the work of Di Persio L, Honchar O and also incorporated technical indicators to predict future prices. Their model achieved an accuracy of 56.3% in classification but did not record the profits from the trades.

  • LSTM, another deep learning algorithm, captures temporal dynamics and thus proves useful for modeling the time series behavior of the stock market according to the research paper by Nelson DMQ, Pereira ACM, De Oliveira RA (2010). Features can be extracted by convolutional layers while temporal and recurrent characteristics can be captured by LSTM layers.

  • Livieris et al. (2021) proposed a multi-input deep learning model for cryptocurrencies, known as MICDL. The proposed method utilizes individual cryptocurrency data as inputs in a convolutional neural network (CNN) layer, followed by a pooling layer and an LSTM layer. Classic structures of deep neural networks, such as dense layers, batch normalization, dropout, and output layers, are also leveraged. The CNN-LSTM architecture achieved an accuracy of 55.03% on Bitcoin data, while the accuracy was 51.51% for Ethereum data and 49.61% for Ripple data.

  • McNally et al. (2018) utilized various deep learning algorithms to predict the price of Bitcoin. In the initial phase, functional patterns were extracted from the data through feature engineering techniques. Experimental results showed that the long short-term memory (LSTM) model achieved the highest accuracy of 52.78%, while the recurrent neural network (RNN) had the lowest accuracy of 5.45%.

  • Saadah et al. (2020) applied machine learning and deep learning methods to predict the prices of Bitcoin, Ethereum, and Ripple. The methods included k-nearest neighbors, support vector machine (SVM), and LSTM. The experimental results demonstrated that LSTM achieved optimal root mean square error (RMSE) for all three types of cryptocurrencies, with RMSE values of 928.62 for Bitcoin, 11.69 for Ethereum, and 0.16 for Ripple.

  • Similar cryptocurrency price predictions were conducted by Sebastiao et al. (2021). The authors devised several machine learning models, including linear regression, random forest, and SVM, to assess the predictive ability of cryptocurrencies. The testing results on Ethereum prices reached a win rate of 63.33% for the strategies. Alternatively, linear models achieved optimal RMSE when forecasting Ethereum and Litecoin prices with values of 6.85 and 8.14, respectively, while random forest attained an RMSE of 5.77 when predicting Bitcoin prices.

  • Arevalo et al. (2016) used deep neural networks with five layers to predict Apple stocks. The method employed by the authors for stock trading relied on the difference between predictions and actual values, which could lead to overtrading. The authors demonstrated the power of deep neural networks (DNNs) compared to shallow networks, as deep networks achieved higher accuracy. Shallow networks were incapable of effectively processing the data, which could result in inaccurate predictions.

PaperPredictive algorithmsDataset

Hoseinzade E, Haratizadeh S (2019)

CNN

S&P500, DJI

Livieris et al. (2021)

CNN + LSTM

Ethereum

Ripple

Saadah et al. (2020)

LSTM

Ethereum

Ripple

Arevalo et al., 2016

DNN

Apple Stock

McNally et al. (2018)

LSTM

Bitcoin

SebastiΓ£o, H.

Linear Models

Ethereum

Litecoin

Random Forest

Ethereum

Litecoin

Last updated