Here is a continuation of my article on everything you need to know about cryptocurrency trading:
Comparison of three cryptocurrency trading systems
Real-time trading systems use real-time functions to collect data and generate trading algorithms. The turtle trading system and the arbitrage trading system show a sharp contrast in the behavior of returns and risks. Using the turtle trading system in the cryptocurrency market can obtain high returns and high risks. Although the arbitrage trading system is inferior in terms of returns, it also has lower risks. A common feature of the Turtle Trading System and Arbitrage Trading System is that they perform well in capturing Alpha.
Many researchers focus on the analysis of technical indicators (patterns) of cryptocurrency market transactions. Research examples of this method include "Turtle Soup pattern strategy", "Nem (XEM) strategy", "Amazing Gann Box strategy", "Busted Double Top Pattern strategy", " Bottom Rotation Trading strategy". Table 6 shows a comparison of these five classic technical trading strategies using technical indicators. "Turtle Soup pattern strategy" uses a 2-day price breakout to predict the price trend of cryptocurrencies. This strategy is a chart trading model. "Nem (XEM) strategy" combines the rate of change (ROC) indicator and the relative strength index (RSI) to predict price trends . "Amazing Gann Box strategy" predicts the exact ups and downs of Gann Box and is used to capture the explosive trend of cryptocurrency prices. Technical analysis tools such as candlesticks and box charts are used on the basis of the Fibonacci Golden Ratio. Fibonacci retracements use horizontal lines to indicate the levels of support and resistance that may exist in the market. The "Bankruptcy Double Top Model" uses a bear market reversal trading model to generate sell signals to predict price trends. "Bottom rotation trading" is a method of technical analysis that selects the bottom before the reversal occurs. This strategy uses price chart patterns and block diagrams as technical analysis tools.
Sungjoo et al. conducted a survey using genetic programming (GP) to discover attractive technological patterns in the cryptocurrency market. In the experiment, 12 technical indicators such as moving average method (MA) and stochastic oscillation method were used, and methods such as adjustment gain, matching count, relative market pressure and diversity measurement were used to quantify the attractiveness of the technical model. Through extended experiments, the GP system successfully found an attractive technical model, which is useful for portfolio optimization. Hudson et al. applied nearly 15,000 technical trading rules (divided into MA rules, filtering rules, support and resistance rules, oscillation rules and channel breakthrough rules). This comprehensive study found that technical trading rules provide investors with significant predictive power and profitability. Corbet et al. analyzed various technical trading rules and trading range breakthrough strategies in the form of moving average oscillators to generate higher returns in the cryptocurrency market. Using one-minute dollar-denominated Bitcoin closing price data, backtesting shows that the variable-length moving average (VMA) rule performs best when it generates the most useful signals in high-frequency trading.
Paired trading is a trading strategy that tries to use the mean reversion between the prices of certain securities. Miroslav used the benchmark of Gatev et al. to investigate the applicability of the standard to transaction methods for cryptocurrency data. The paired trading strategy is constructed in two steps. First, determine the right pair with a stable long-term relationship. Second, calculate the long-term equilibrium, and determine the paired trading strategy based on the spread. The study also expanded intraday trading pairs using high-frequency data. In general, in Miroslav's experiment, the model can achieve a monthly profit of 3%. Broek applied co-integration-based pair transactions in cryptocurrency transactions and found that 31 pairs of pair transactions were significantly co-integrated (intra-department and cross-department). By selecting four pairs and testing them within a 60-day trading cycle, the paired trading strategy has gained profitability from arbitrage opportunities, thus negating the effective market hypothesis (EMH) of the cryptocurrency market. Lintihac et al. proposed an optimal dynamic matching trading strategy model for asset portfolios. The experiment uses stochastic control techniques to calculate the optimal portfolio weights and correlates the results with other strategies commonly used by practitioners (including static dual-threshold strategies). Thomas et al. proposed a pairwise trading model that combines time-varying volatility with constant elasticity of variance. The experiment uses the finite difference method to calculate the best pairing strategy, and uses the generalized moment method to estimate the parameters.
Other system trading methods in cryptocurrency trading mainly include informed trading. Using USD/Bitcoin exchange rate transaction data, Feng et al. found evidence of informed transactions in the Bitcoin market. Compared with the quantile of seller-initiated (buyer-initiated) orders, the quantile of buyer-initiated (seller-initiated) orders The number is unusually high before a large positive (negative) event; this study uses a new indicator inspired by the volume imbalance indicator. Evidence of informed transactions in the Bitcoin market shows that investors profit from their private information before they obtain it.
Emerging crypto trading technology
Copula-quantile causality analysis and Granger causality analysis are two methods for studying causality in cryptocurrency transaction analysis. Bouri et al. applied the Copula-quantile causality method to the volatility of the cryptocurrency market. The experimental method extends the Copula-Granger distribution causality (CGCD) method proposed by Lee and Yang in 2014. The experiment constructs two CGCD tests with the copula function. The parameter test uses the six-parameter copula function to find the dependence density between variables. The performance matrix of these functions varies with the copula density. The focus of the study is on three distribution areas: left tail (1%, 5%, 10% quantile), central area (40%, 60% quantile and median) and right tail (90%, 95% , 99% quantile). The research provides important evidence of Granger causality from transaction volume to the left and right tail returns of the seven large cryptocurrencies. Elie et al. tested the causal relationship between the volatility of major cryptocurrencies through the frequency domain of Bodart and Canon, and distinguished between temporary and permanent causality. The results show that in the long run, permanent shocks are more important to explain Granger's causality, while short-term shocks dominate the causality of small cryptocurrencies. Badenhorst tried to use Granger causality method and ARCH (1,1) to reveal whether the trading volume of spot and derivatives market affects Bitcoin price fluctuations. The research results show that spot trading volume has a significant positive impact on price fluctuations, while the relationship between cryptocurrency fluctuations and the derivatives market is uncertain. Elie et al. used the Dynamic Equivalent Correlation (DECO) model and reported evidence that the average return equilibrium correlation between 12 major cryptocurrencies changes over time. The results show that despite the sharp drop in cryptocurrency prices in 2018, the degree of integration of the cryptocurrency market has increased. In addition, the measurement and uncertainty of transaction volume are key determinants of integration.
Some econometric methods in time series research, such as GARCH and BEKK, have been used in the cryptocurrency trading literature. Conrad et al. used the GARCH-MIDAS model to extract the long-term and short-term volatility components of the Bitcoin market. The technical details of the model decompose the conditional variance into low-frequency components and high-frequency components. The study found that the realized volatility of the S&P 500 index has a significant negative impact on the long-term bitcoin volatility, and the volatility risk premium of the S&P 500 index has a significant positive impact on the long-term bitcoin volatility. Ardia et al. used the Markov Transformation GARCH (MSGARCH) model to test whether there are institutional changes in the GARCH volatility dynamics of Bitcoin's logarithmic return. In addition, Bayesian method is used to estimate model parameters and calculate VaR prediction. The results show that the MSGARCH model is significantly better than the single mechanism GARCH model in terms of value-at-risk prediction. Troster et al. conducted general GARCH and GAS (Generalized Autoregressive Score) analysis to model and predict the benefits and risks of Bitcoin. Experiments have found that the heavy-tailed GAS model can provide the best out-of-sample prediction and goodness-of-fit attributes for Bitcoin's return and risk modeling. The results of the study also illustrate the importance of modeling excessive kurtosis of Bitcoin returns. Charles et al. studied four cryptocurrency markets, including Bitcoin, Dash, Litecoin, and Ripple. The results show that, in addition to the Dash market, the salient feature of cryptocurrency yields is the existence of jumps and structural breakthroughs. Four GARCH models (namely GARCH, APARCH, IGARCH and FIGARCH) and three types of income with structural mutations (original income, jump filtered income and jump filtered income) are considered. This research demonstrates the importance of cryptocurrency volatility jumps and structural breakthroughs.
Some researchers focus on long memory methods for the volatility of the cryptocurrency market. The long memory method focuses on the long-term correlation between market fluctuations and the significant long-term correlation. Chaim et al. estimated a multivariate random volatility model with discontinuous jumps in the cryptocurrency market. The results show that long-term volatility seems to be driven by major market developments and general interest rates. Caporale et al.  tested the durability of the cryptocurrency market through rescaled interval (R/S) analysis and score points. The research results show that the market is continuous (there is a positive correlation between its past and future values), and its level changes over time. Khuntin et al.  applied the adaptive market hypothesis (AMH) to the predictability of Bitcoin returns. The consistency test of Dominguez and Lobato , and the generalized spectrum (GS) of Escanciano and Velasco  are used to capture the time-varying linear and nonlinear dependence of Bitcoin's return. The research results verify that the evidence of evolutionary efficiency and dynamic efficiency in Bitcoin price changes is consistent with AMH's statement.
Katsiampa et al.  applied three pairs of bivariate BEKK models in 2018 to test the dynamics of conditional volatility and the interconnection and conditional correlation between three pairs of cryptocurrencies. More specifically, the BEKK-MGARCH method also captures the cross-market effects of shocks and volatility, which are also called shock conduction effects and volatility spillover effects. The experiment found evidence of the two-way shock propagation effect between Bitcoin and Ethereum and Litcoin. In particular, there is a two-way shock spillover effect between Bitcoin, Ethereum and Litcoin, and there is a time-varying condition correlation, and a positive correlation is dominant. In 2019, Katsiampa  further studied an asymmetric diagonal BEKK model to test the conditional variance of five cryptocurrencies, which are significantly affected by previous square errors and past conditional volatility. The experiment tested the null hypothesis and stationarity hypothesis of unit root. Under the premise of ensuring stationarity, the ARCH effect test is performed on ARCH-LM to test the requirements of volatility modeling of the return rate series. In addition, this article also uses a multivariate GARCH model to test the volatility coordination between cryptocurrency pairs. The results confirmed the non-normality and heteroscedasticity of price returns in the cryptocurrency market. This finding also determines the impact of cryptocurrency volatility dynamics due to major news. Hultman  set out to study GARCH (1,1), bivariate BEKK (1,1) and a standard random model to predict Bitcoin's volatility. The rolling window method was used in the experiment. Mean absolute error (MAE), mean square error (MSE) and root mean square deviation (RMSE) are three loss criteria to evaluate the degree of error between the predicted value and the true value. The results show that for the three different loss criteria, the order of the loss function is: GARCH(1,1)>binary BEKK(1,1)>standard random; in other words, GARCH(1,1) is predicting Bitcoin The best performance in terms of volatility. Wavelet time-scale persistence analysis has also been applied to the prediction and research of the volatility of the cryptocurrency market . The results show that the information efficiency and volatility persistence of the cryptocurrency market are highly sensitive to time scales, yield and volatility measures, and institutional changes. Adjepong et al.  are linked to similar studies by Corbet et al.  and show that GARCH absorbs new information about data faster than BEKK.
Machine Learning Technology
As mentioned earlier, machine learning technology constructs computer algorithms to automatically improve itself by finding patterns in existing data without explicit instructions . The rapid development of machine learning in recent years has promoted its application in cryptocurrency transactions, especially in the prediction of cryptocurrency revenue.
This article reviews commonly used machine learning techniques
Several machine learning techniques are used in cryptocurrency transactions. We distinguish these algorithms by target set: classification, clustering, regression, and reinforcement learning. Due to the inherent changes and widespread adoption of deep learning technology, we have divided a section to discuss deep learning.
Classification algorithm. The goal of classification in machine learning is to classify input objects into different categories as needed, and we can assign labels to each category (for example, upper and lower). Based on the literature we collected, Naive Bayes (NB) , Support Vector Machine (SVM) , K Nearest Neighbor (KNN) , Decision Tree (DT) , Random Forest (RF) )  and Gradient Boosting (GB)  algorithms have been used in cryptocurrency transactions. NB is a probabilistic classifier based on Bayes' theorem, with strong (naive) conditional independence assumptions between features . Support vector machine (SVM) is a supervised learning model whose purpose is to implement a high-margin classifier in combination with learning boundary theory . The support vector machine assigns new examples to one or the other category, making it a non-probabilistic binary linear classifier , although some modifications can be used to interpret its output probabilistically . KNN is an algorithm based on memory or delayed learning, in which functions are only locally approximated, and all calculations are postponed to the inference time . DT is a decision support tool algorithm that uses a tree-like decision diagram or model to divide the input pattern into multiple regions, and then assigns a relevant label to each region . RF is an integrated learning method. This algorithm constructs a large number of decision trees in the training process, outputs the average consistency as the prediction class in the case of classification, and outputs the average prediction value in the case of regression . GB generates prediction models in the form of a set of weak prediction models .
Clustering Algorithm. Clustering is a machine learning technique that groups data points so that each grouping shows a certain regularity . K-Means is a vector quantization method for cluster analysis in data mining. K-means storage is used to define the centroid of the cluster; if a point is closer to the centroid of the cluster than any other centroid, it is considered to be in a particular cluster . According to the literature we collected, K-Means is one of the most commonly used clustering algorithms in cryptocurrency trading.
Regression algorithm. We define regression as any statistical technique designed to estimate continuous values . Linear regression (LR) and scatter plot smoothing are commonly used techniques to solve regression problems in cryptocurrency trading. LR is a linear method used to simulate the relationship between a scalar response (or dependent variable) and one or more explanatory variables (or independent variables) . Scatter plot smoothing is a technique for fitting functions through scatter plots to best represent the relationship between variables .
Deep learning algorithm. Deep learning is a modern form of artificial neural network (ANN) , which is made possible due to advances in computing power. An artificial neural network is a computing system inspired by the natural neural networks that make up animal brains. The system "learns" to perform tasks including predictions by considering examples. The superior accuracy of deep learning comes from high computational complexity and high cost. Deep learning algorithms are currently the basis of many modern artificial intelligence applications . Convolutional Neural Networks (CNNs) , Recurrent Neural Networks (RNNs) , Gated Recurrent Units (GRU) , Multilayer Perceptrons (MLP) and Long-Short-Term Memory (LSTM)  The network is the most commonly used deep learning technology in cryptocurrency transactions. CNN is a special type of neural network layer, usually used for supervised learning. CNNs have achieved the greatest success in image processing and natural language processing. See  for an attempt to use CNN in cryptocurrency. RNN is an artificial neural network in which the connections between nodes form a directed graph with possible cycles. Due to the introduction of memory in the loop connection, this structure of RNN makes it suitable for processing time series data . However, they face the vanishing gradient problem , so different changes have recently been proposed. LSTM  is a special RNN architecture that is widely used. In financial time series, LSTMs have been shown to be superior to non-hierarchical RNNs because of their ability to selectively remember patterns for a long time. GRU  is another gated version of the standard RNN, which has been used for encrypted transactions . Another deep learning technique used in cryptocurrency transactions is Seq2seq, which is a concrete implementation of the encoder-decoder architecture . The original purpose of Seq2seq was to solve the problem of natural language processing, but it was also applied to cryptocurrency trend prediction in .
Reinforcement learning algorithm. Reinforcement learning (RL) is a field of machine learning that uses the behavior of software agents in the environment to maximize cumulative returns . Deep Q Learning (DQN)  and Deep Boltzmann Machine (DBM)  are commonly used techniques in cryptocurrency transactions using RL. Deep Q learning uses neural networks to approximate the Q value function. The state is used as input, and the Q value of all possible actions is generated as output . DBM is a binary paired Markov random field (undirected probability graphical model), with multiple hidden random variables . It is a randomly coupled random binary unit network.
Research on machine learning models
In the development of machine learning trading signals, technical indicators are often used as input features. Nakano et al.  studied Bitcoin intraday technical transactions based on ANNs for revenue prediction. The experiment obtained Bitcoin's mid-frequency price and transaction volume data from cryptocurrency exchanges (the data interval is 15 minutes). The artificial neural network predicts the price trend (up and down) of the next period based on the input data. The data is preprocessed to construct a training data set containing a matrix of technical patterns such as EMA, emerging market small-cap stocks (EMSD), and relative strength index (RSI). The numerical experiments include different research aspects, including basic ANN research, different Level effects, effects of different activation functions, different outputs, different inputs and effects of additional technical indicators. The research results show that compared with the original technical trading strategy, the use of various technical indicators may prevent overfitting in the classification of non-stationary financial time series data, thereby improving trading performance. (Buy and hold is the benchmark strategy for this experiment.)
By predicting price trends, the classification regression machine learning model is applied to cryptocurrency trading. Most researchers focus on the comparison of different classification and regression machine learning methods. Sun et al.  used random forests (RFs) and factors in Alpha01  (obtaining features from the history of the cryptocurrency market) to build predictive models. The experiment collects data from the API of the cryptocurrency exchange, and selects 5 minutes of frequency data for backtesting. The results show that performance is proportional to the amount of data (the more data, the higher the accuracy), and the factors used in the RF model seem to have different importance. For example, the "Alpha024" and "Alpha032" features are the most important in the model used. (The alpha feature comes from the paper "101 Formulaic Al phas" ) Vo et al.  applied RFs to high-frequency cryptocurrency trading (HFT) and compared them with deep learning models. When using forward fill imputation method to replace empty values (that is, missing values), collect minute-level data. Different periods and RF trees were tested in the experiment. The author also compared the F1 accuracy and recall metrics of RF and deep learning (DL). The results show that despite the presence of multicollinearity in ML features, RF is still effective, but the lack of model recognition may also lead to model recognition problems; this study also tried to create a Bitcoin HFT strategy using RF. Maryna et al.  studied the profitability of algorithmic trading strategies based on training SVM models to identify cryptocurrencies with high or low predicted returns. The results show that the performance of the support vector machine strategy ranked fourth, only better than the S&P which simply bought and held the S&P index. pbh strategy. (There are four other benchmark strategies in this study) The author observed that support vector machines require a large number of parameters, so it is prone to overfitting, resulting in poor performance. Barnwal et al.  used productions and discriminative classifiers to establish a superposition model, in particular, 3 productions and 6 discriminative classifiers were combined through a single-layer neural network to predict the price trend of cryptocurrencies. The discriminative classifier directly models the relationship between the unknown data and the known data, while the generative classifier indirectly models the prediction through the data generation distribution . Technical indicators include trend, momentum, volume and volatility as the characteristics of the model. The author discussed the impact of different classifiers and features on prediction. Attanasio et al.  compared various classification algorithms, including SVM, NB, and RF, to predict the next-day price trend of a given cryptocurrency. The results show that due to the heterogeneity and volatility of cryptocurrency financial instruments, a prediction model based on a series of predictions is superior to a single classification technique in cryptocurrency transactions. Madan et al.  modeled the Bitcoin price prediction problem as a binomial classification task, using a custom algorithm using random forests and generalized linear models to conduct experiments. The experiment uses daily data, 10-minute data, and 10-second data. Experiments show that 10-minute data has better sensitivity and specificity than 10-second data (10-second prediction accuracy is about 10%). Considering predictive trading, 10 minutes of data helps to show clearer trends in the experiment compared to a 10-second backtest. Similarly, Virk  compares RF, SVM, GB and LR to predict the price of Bitcoin. The results show that among the binomial classification machine learning algorithms, the support vector machine has the highest classification accuracy of 62.31%, and the classification accuracy is 0.77.
Different deep learning models have been used to find price movement patterns in the cryptocurrency market. Zhengy et al.  implemented two machine learning models, namely fully connected ANN and LSTM to predict the price dynamics of cryptocurrencies. The results show that although LSTM is theoretically more suitable for time series dynamic modeling than ANN, ANN is generally better than LSTM; in joint forecasting (five cryptocurrency daily price forecasts), the performance indicators considered are MAE and RMSE. The research results show that the future state of the cryptocurrency time series depends largely on its historical evolution. Kwon et al.  used the LSTM model with a three-dimensional price tensor representing the past price changes of cryptocurrencies as input. This model is superior to the GB model in terms of F1 performance. Specifically, in the 10-minute price prediction, its performance is about 7% higher than the GB model. In particular, experiments show that LSTM is more suitable for classifying cryptocurrency data with high volatility. Alessandretti et al.  tested the prediction of daily cryptocurrency prices by gradient-enhanced decision trees (including single regression and XGBoost enhanced regression) and LSTM models. They found that the method based on gradient boosting decision tree works best when predicting based on a short-term window of 5/10 days, while LSTM works best when predicting based on 50 days of data. The relative importance of the features in the two models is compared, and the optimal combination based on geometric mean return and Sharpe ratio is discussed. Phaladisailoed et al.  chose regression models (Theil-Sen regression and Huber regression) and deep learning-based models (LSTM and GRU) to compare the performance of predicting the rise and fall of bitcoin prices. Among the two commonly used metrics, MSE and RSquare (R2), GRU has the highest accuracy. Fan et al.  applied an LSTM structure enhanced by an autoencoder when predicting the median price of 8 pairs of cryptocurrencies. ^^ The real-time data of the 2-level limit order book was collected, and the experiment achieved a 78% accuracy rate of price change prediction in high-frequency trading (tick-level). ^^ This research improves and verifies the view of Sirignano et al.  that for the cryptocurrency market, the general model has better performance than the currency pair specific model. In addition, “walk-through” (that is, retraining the original deep learning model when it seems no longer effective) is also proposed as a method to optimize the training of deep learning models, and has shown a significant improvement.
To be continued.......