中文版

因子流：

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

利用11個數字貨幣的行情和發行資料，構造了動量、價值和caryy因子：

動量因子指標是

動量因子為前一週的回報

價值因子為當前市場價值與區塊鏈中$-value鏈上交易的7天平均值的比值。

Carry因子被定義為7天內發行數字貨幣總量的負數，除以7天週期期發行的數字貨幣

三個因子都有效，動量因子表現最好，因子複合後能獲得更高的風險調整後收益。

演算法流：

CONTROL STRATEGY TO TRADE CRYPTOCURRENCIES

JOSEF KOKEŠ

Main goal of this article is to introduce strategy for automated trading on cryptocurrency exchange market. For this purpose we will use algorithm based of Floyd-Warshall algorithm

這篇文章的主要目的是介紹在數字貨幣交易市場下使用演算法進行自動交易，介紹的演算法基於的是最短路徑弗洛伊德演算法，這種演算法的時間複雜度為O(N^3)，空間複雜度為O(N^2)

然而並沒有給出比較有價值的內容，可以使用HDOJ 1217 Arbitrage(擬最短路，floyd演算法)裡面的題目和程式碼進行測試，可以使用floyd演算法和SPFA演算法進行測試

對於預測模型呢，

多因子選股策略主要步驟是因子池構建、檢驗因子有效性、剔除冗餘因子、資料預處理、分類模型的構建、模型準確率及歷史回測評估。

Tips：一般來說冗餘因子剔除的方法如下：首先根據不同因子與模型形成期的股票

投資組合的收益率的相關性，對不同的因子的組合打分；接下來按照單隻股票計

算各因子之間的得分相關係數矩陣；根據上述計算的得分相關係數矩陣計算整個

形成期的相關係數矩陣的平均值；最後設定一個選擇閾值，用於剔除冗餘因子。

技術面指標如動量、換手率、波動率等和其他指標如預期收益增長、巨集觀經濟變數

等。

《Automated Bitcoin Trading via Machine Learning Algorithms》

使用xgboost具有自動隨機選擇因子的能力，可以取代在較多或者大量因子模型的因子有效性檢驗的問題。

作者：斯坦福大學 計算機學院的Isaac Madan Shauraya Saluja

特徵：使用了過去5年的和比特幣價格和支付相關的25個特徵，

資料：使用了兩種時間段，每日的記錄，時間間隔為10分鐘

結果：我們能夠以98.7%的精確度預測出每日價格變化的跡象。

模型的假設是未來的價格可以看作是過去價格序列的組合，並把問題看作一個二分類問題，使用了模型有隨機森林和廣義線性模型（generalized linear models.）

先前的工作：Shah and Zhang使用了貝葉斯迴歸（Bayesian regression）獲得了較好的收益。

他們的不足：沒有探索比特幣價格和其市值，挖礦速度上面的關係

結論：

第一個結論：

使用了廣義線性迴歸模型，結果很不錯，資料集價格點之間的較長時間間隔可能是導致價格波動不準的原因，使用SVM進行分類的效果比較差，原因可能是因為資料量不足的問題？

對隨機森林而言，相對於廣義二分線性迴歸模型在資料集上表現出了更高的精準度但是預測性不如廣義線性迴歸模型。較低的預測是因為假陽性情況比較多，比實際情況表現出了更positive。

第二個結論：

       10分鐘級的資料比10s級的資料效果更好一些，表現在靈敏度和特異性比率，比10s級的資料更能反映趨勢。RF相對於二分線性迴歸效果好的原因可能是因為RF使用非引數決策樹，所以資料的離群值和線性可分性不受關注。

英文版：

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

This paper uses momentum and distribution data from 11 digital currencies to construct momentum, value, and caryy factors.

The momentum is defined as prior week’s return

The value is defined as the ratio of current market value and the trailing 7 days average of $-valued on-chain transactions in its blockchain.

Carry is defined as negative of the sum total coin issuance over the precding 7days,divided by the coins outstanding at the beginning of that 7 day period

momentum as “the return we get if things keep changing the way they have”, value as the return we get “if things go back to where they were/some kind of fair equilibrium”, and carry as “the return we get if things don’t change at all”.

All three factors are effective, the momentum factor performs best, and the factor composite can obtain higher risk-adjusted returns.

《Automated Bitcoin Trading via Machine Learning Algorithms》

hypothesis

Specifically, with the idea that future price trends can be inferred directly from a linear combination of existing time series data

Dataset

       Our data set consists of over 25 features relating to the Bitcoin price and payment network over the course of five years, recorded daily.

Result

Using this information we were able to predict the sign of the daily price change with an accuracy of 98.7%.

論文《Cryptocurrency price drivers: Wavelet coherence analysis revisited》（小波相干性分析）

 google引用量 38

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

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

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》

如題目所示，流行病