Resumo

Cryptocurrencies were firstly introduced by Nakamoto (2009), that proposed a digital currency and payment system based on cryptographic proof instead of the trust of third parties such as financial institutions. This electronic cash was named Bitcoin and still stands as the main cryptocurrency in the market. However, since the creation of Bitcoin, a big number of cryptocurrencies have been created and gained a significant share of this emerging market.

Cryptocurrency market prices are marked by a high volatility, presenting opportunities to investors searching for higher returns and diversification. This gives an important role in predicting its futures prices to enable trading strategies and providing information to investors. This paper aims to discuss the implementation and prior results obtained on the one-step prediction of cryptocurrencies’ closing prices on a daily frequency using Artificial Neural Networks (ANN), comparing their predictive performance against state of the art time series econometric models.

ANNs have been intensively applied in the financial market due to their ability to capture non-linearities. Bidirectional LSTM (BiLSTM), primarily used on natural language processing, distinguishes itself by considering the influence of future factors on the present, so the time series are trained in both directions. Varied features can be used as input to those models. Google Trends data, which presents the relative volume of queries on Google search engine over time, presents a promising alternative as it can be used to anticipate future movements on an asset (Preis et al., 2013).

ANN algorithms were implemented to one-step prediction of 10 most traded cryptocurrencies: Bitcoin, Ethereum, Binance Coin, Cardano, Ripple, Dogecoin, Litecoin, Chainlink, Tron and Stellar. The results obtained by Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM) and BiLSTM were compared with results of state of the art models, Autoregressive Integrated Moving Average (ARIMA) and Exponential Smoothing (ETS), and with a Random Walk implementation. Previous exchange rate is used as input to all the models. A BiLSTM approach is tested combining Google Trends data to inputs.

Besides the good results from the statistical methods, especially the ETS, the BiLSTM model enhanced with Google Trends data surpassed all the other methods by a large margin in terms of directional accuracy and presented smaller values for absolute and squared errors for part of the samples. For samples of Binance Coin, Ethereum, Chainlink and Litecoin, it achieved a directional accuracy of over 70%, while no accuracy was lower than 57% for all the currencies.

Traditional statistical methods lead to good and consistent results, showing why these methods are considered state of the art on time series prediction. The implemented BiLSTM surpasses the traditional LSTM, comproving it is a good fit to financial time series. Can also be concluded that Google Trends data can be successfully added to the previous exchange rates as input on forecasting cryptocurrencies' closing prices leading to significant improvements in the accuracy of the forecasts.

Nakamoto, S. (2009). Bitcoin: A peer-to-peer electronic cash system. Nakano, M., Takahashi, A., & Takahashi, S. (2018). Bitcoin technical trading with artificial neural network. Physica A: Statistical Mechanics and its Applications, 510, 587–609. Preis, T., Moat, H. S., & Stanley, H. E. (2013). Quantifying trading behavior in financial markets using google trends. Scientific Reports, 3(1). Yang, M., & Wang, J. (2022). Adaptability of financial time series prediction based on bilstm. Procedia Computer Science, 199, 18–25.