数字货币量化投资文献综述（二）

《“Know When to Hodl 'Em, Know When to Fodl 'Em”: An Investigation of Factor Based Investing in the Cryptocurrency Space》

数据特征：

利用11个数字货币的行情和发行数据，构造了动量、价值和caryy因子：

动量因子为前一周的回报

价值因子为当前市场价值与区块链中$-value链上交易的7天平均值的比值。

Carry因子被定义为7天内发行数字货币总量的负数，除以7天周期期发行的数字货币

结果：三个因子都有效，动量因子表现最好，因子复合后能获得更高的风险调整后收益。

《Bayesian regression and Bitcoin》

       Dataset：

 

《Automated Bitcoin Trading via Machine Learning Algorithms》

数据特征：使用了两种时间间隔的数据，分别为10分钟和10秒级一共过去5年的数据

使用了比特币价格和支付相关的25个特征

结果：我们能够以98.7%的精确度预测出每日价格变化的迹象。

模型的假设是未来的价格可以看作是过去价格序列的组合，并把问题看作一个二分类问题，使用了模型有随机森林和广义线性模型（generalized linear models.）

结论：

(1)使用了广义线性回归模型，结果很不错，数据集价格点之间的较长时间间隔可能是导致价格波动不准的原因，使用SVM进行分类的效果比较差，原因可能是因为数据量不足的问题？

对随机森林而言，相对于广义二分线性回归模型在数据集上表现出了更高的精准度但是预测性不如广义线性回归模型。较低的预测是因为假阳性情况比较多，比实际情况表现出了更positive。

（2）10分钟级的数据比10s级的数据效果更好一些，表现在灵敏度和特异性比率，比10s级的数据更能反映趋势。RF相对于二分线性回归效果好的原因可能是因为RF使用非参数决策树，所以数据的离群值和线性可分性不受关注。

《“Know When to Hodl 'Em, Know When to Fodl 'Em”: An Investigation of Factor Based Investing in the Cryptocurrency Space》

Data factors:

Momentum, value and caryy factor are constructed by using the market quotation and issuing data of 11 digital currencies:

Momentum factor is the return of the previous week

The value factor is the ratio between the current market value and the 7-day average value of transactions on the $-value chain in the block chain.

Carry factor is defined as the negative number of the total amount of digital currency issued within 7 days, divided by the amount of digital currency issued during the 7-day period

Results: all three factors were effective, momentum factors performed best, and the combination of these factors could achieve higher risk-adjusted returns.

 

《Automated Bitcoin Trading via Machine Learning Algorithms》

Data factors: data at two time intervals, 10 minutes and 10 seconds, being used for a total of 5 years

25 features related to bitcoin price and payment are used

The result: the model was able to predict the signs of daily price changes with 98.7% accuracy.

The model assumes that the price in the future can be seen as a combination of the price sequences in the past, and the problem can be considered as a binary classification problem. The model is generalized linear model with a random forest.

Conclusion:

(1) the generalized linear regression model is used, and the result is very good. The long time interval between the price points of the data set may be the cause of inaccurate price fluctuations, and the classification effect using SVM is relatively poor.

For the random forest, compared with the generalized bipartite linear regression model, the accuracy of the data set is higher than that of the generalized bipartite linear regression model, but the predictability is not as good as that of the generalized linear regression model.The lower prediction is because there are more false positives and more positive than the actual situation.

(2) the 10-minute data is more effective than the 10s data, which is reflected in the sensitivity and specificity ratio and can reflect the trend better than the 10s data. The reason why RF performs better than bipartite linear regression may be that RF USES non-parametric decision trees, so the outliers and linear separability of data are not concerned.

 

《Cryptocurrency price drivers: Wavelet coherence analysis revisited》

Hypothesis：The hypothesis of this paper is that the relationship between online factors and prices depends on market mechanisms.

Method：

Using wavelet coherence to study the co-movement between the cryptocurrency price and its related factors

Result：

The main finding of this study is that the medium-term positive correlation between factors extracted from the Internet and prices is significantly enhanced when the price series is foamy; this explains why these relationships appear and disappear over time. The second finding is that the short-term relationship between the chosen factors and the price seems to be caused by specific market events (such as hacking/security vulnerabilities), and the impact of these factors on prices is not consistent over time intervals. In addition, the relationship between different cryptocurrencies was studied for the first time using wavelet coherence.

 

《Predicting Cryptocurrency Price Bubbles Using Social Media Data and Epidemic Modelling》

It is our hypothesis that it is possible to examine patterns in social media usage to detect the earlier stages of a cryptocurrency price bubble, the boom phase referenced above and the 'increasing interest' described by Shiller.

 

The link between cryptocurrency prices and social media usage has already been demonstrated in the literature.

Dataset:

The subreddit on Reddit, there are four factors:

Posts, Subscriber growth, New authors, volume.

Tips：The profit-depressing effect of volume is because volume tends to lag social media usage in moving to an epidemic state, and therefore positions are entered later

Module:

HMM AIR (two states, epidemic and nonepidemic) to detect bubble-like behaviors in the time series

Conclusion: that social media data can play an important role in forecasting cryptocurrency movements.

 

《Predictive Analysis of Cryptocurrency Price Using Deep Learning》

This paper presents a approach to predicting cryptocurrency prices by considering various factors such as market capitalization, volume, cyclic supply, and maximum supply

using deep learning techniques such as recurrent neural network (RNN) and long-term short-term memory (LSTM). The experimental results verify the applicability of this method in the accurate prediction of cryptocurrency prices. The results of the study indicate that market open may play a key role in influencing all other parameters. In addition, the size of the data set may affect future predictions, as the results of models trained with large data sets perform better.

《Deep direct reinforcement learning for financial signal representation and trading》

They proposed a deep direct reinforcement learning framework for financial signal representation and trading. They focus on training computers to outperform experienced financial traders in predicting the exact results of financial transactions. They proposed combining reinforcement learning (RL), deep learning (DL) and recursive deep neural network (NN) to obtain accurate prediction results. They used the commodity futures market and stock market data to verify the proposed approach.

《Investigating Factors behind Choosing a Cryptocurrency》

They looked at two types of problems related to cryptocurrencies: exploring various factors that affect users' use of mine cryptocurrencies, as well as factors that affect the popularity and value of cryptocurrencies. They explored eight kinds of cryptocurrencies and constructed a survey. more than half of the participants believe that the currency name and logo affect the choice of using and/or mining a cryptocurrency. The ease of mining, community, anonymity, privacy is also as advantages of the currencies people choose to use and/or mine.

《Neural network approximation precision change analysis on cryptocurrency price prediction》

In this paper, we can see that Increasing the market data input of relevant parameters can improve the accuracy of the prediction. The methods described in Chapter 5 based on Fourier analysis can be used with different data sets to provide additional input parameters. And in the experiments, the LSTM network was slightly more accurate than the MLP network.