Cryptocurrency prices are notoriously unpredictable. Their value can fluctuate based on news, popularity and court rulings; recent favorable decisions seem to have given XRP a boost but its lasting effect remains uncertain.
This paper assesses the predictive ability of machine learning models on bitcoin, ethereum, and litecoin using model ensembles. Furthermore, this investigation explores their dynamic interactions with volatility indices under various market conditions.
Time series analysis
Cryptocurrency prices have witnessed dramatic price surges since early 2017, with trading volumes skyrocketing exponentially; yet liquidity connectiveness between cryptocurrency assets has remained relatively low compared to traditional securities markets. This study investigates time-varying network interactions such as liquidity connectedness, risk diffusion and contagion among cryptocurrency returns using cross-quantilogram method and quantile connectedness approach in order to explore tail dependence between five cryptocurrencies (BTC, ETH, LTC XRP) and three CBOE uncertainty indices (BTC ETH LTC XRP).
The results demonstrate that cryptocurrency are net pairwise transmitters of system shocks with the exception of XRP which acts as a receiver. Furthermore, their relative magnitudes depend on quantile: an upper quantile may exhibit more complex connectedness networks than median and lower quantiles while lower quantiles tend to be simpler than median ones; also innovation flows depend on quantile; positive spillover skew values on lower quantiles are observed while negative ones on upper ones.
Linear regression model
A linear regression model is an accessible and popular form of predictive analysis that seeks to predict the value of an outcome variable using independent variables as predictors. Furthermore, the model can also help us understand how each independent variable affects its dependent variable; to do this effectively however, samples must be large enough so as not to overfitting the model and overstretch its capabilities.
The least-squares method is an effective algorithm for fitting linear models. This technique calculates the best line by minimizing sum of squared vertical deviations of data points from their linear path; also known as ordinary least-squares (OLS) method.
Cryptocurrencies have seen explosive growth this year, yet their volatility has made some investors wary. Bitcoin and Litecoin remain popular choices; however, over 1,300 other digital currencies exist as well – some more secure than others like Bitcoin which has high systemic risk as only 10,000 addresses control a third of its supply.
LSTM network
LSTM neural networks are recurrent neural networks capable of modeling complex sequential data. They excel at natural language processing tasks like machine translation and text summarization; unlike simple RNNs which suffer from the vanishing gradient problem, LSTMs can learn long-term dependencies among sequences of input data.
Utilizing a special structure called a gate, LSTMs can learn to process words as whole sections rather than word by word, providing more meaningful context capture while helping with solving the vanishing gradient problem by remembering more information from before.
Each LSTM cell features three gates: an input gate, forget gate and output gate. The input gate determines what information enters the cell; forget gates control which information is remembered at each time step; while output gates enable updated information to pass into future time steps.
Time regimes approach
The Time Regimes Approach is a forecasting technique used for cryptocurrency market dynamics by breaking them up into shorter, overlapping sequences. This creates an easier-to-model time series and can help identify an ideal model for each cryptocurrency; additionally, Point Forecast RMSE and Density Forecast CRPS measurements are measured to assess how accurately their models capture volatility of these series.
Table 2 displays the percentage of actual observations outside the 95% credible interval retrieved through simulation, along with each model’s ratio of root mean square error to benchmark RMSE values. As anticipated, using a student-t distribution and including cryptopredictor variables tends to deliver improved performance.
Point forecasting analysis indicates that only the BVAR model outshone its benchmark, likely due to Bitcoin’s lower volatility relative to other cryptocurrencies. While BVAR-GARCH models perform slightly worse for Ethereum and Ripple respectively than VAR, they still outperformed SV models.
