监测加密数字货币操纵行为的机器学习模型

A Machine Learning Model for Monitoring Cryptocurrency Manipulation

加密数字货币匿名性、去中心化、跨区域、全球“7*24”小时交易、监管缺乏等特性促使其市场操纵行为频发。同时,匿名网络社交平台和自动化交易机器人的盛行使得抬价出货操纵行为公开化和短暂化,传统监管手段已无法适用于加密数字货币市场。首先,采用无监督学习模型对加密数字货币交易所分钟级的价量数据进行建模,对异常的量价数据进行快速的监测和预警。然后通过网络实时爬虫技术获取网络社交平台中抬价出货操纵行为的文本信息,与交易所秒级订单数据匹配构建训练集和测试集,采用有监督学习模型对抬价出货操纵行为进行事后识别。模型结果显示:无监督学习模型监测抬价出货的准确率高达84.07%。有监督学习模型在测试集上的精准率和召回率分别为82%和93%,且有监督学习模型的AUC曲线(Area Under Roc Curve)得分为0.83。两种模型均为加密数字货币交易所监管抬价出货行为提供了相关参考。

The anonymity,decentralization,cross-regional,global“7*24”trading and lack of regulation of cryptocurrencies have contributed to the frequent occurrence of market manipulation.At the same time,the prevalence of anonymous online social platforms and automated trading bots has made price manipulation open and transient,and traditional regulatory tools are no longer applicable to the cryptocurrency market.First,unsupervised learning models are used to model the minutelevel price and volume data of cryptocurrency exchanges to quickly monitor and alert abnormal volume and price data.Then,the textual information of price and shipment manipulation behaviors in online social networking platforms is obtained through real-time web crawling technology,and matched with the second-level order data of the exchange to build a training set and a test set,and a supervised learning model is used to identify the price and shipment manipulation behaviors afterwards.The model results show that the unsupervised learning model has an accuracy rate of 84.07%in monitoring the price hikes.The accuracy and recall rates of the supervised learning model on the test set are 82%and 93%,respectively,and the AUC curve(Area Under Roc Curve)of the supervised learning model is 0.83.Both models provide a relevant reference for cryptocurrency exchanges to monitor the bid-rigging behavior.