多代理交易仿真平台中用户交易行为分析的支撑技术研究

发布时间：2018-04-21 08:57

本文选题：多代理系统 + 非阻塞式交互　；参考：《上海交通大学》2015年硕士论文

【摘要】：对于错综复杂的金融交易市场如中国股指期货市场来说,市场价格波动和投资者之间的关系,如投机交易,对于理解市场交易规律和政策设计尤为重要。但是具有不同交易特征的交易者和他们的交易模式对市场的影响仍然不甚明了,这使我们可以基于代理建模(Agent-Based Modeling)设计实现模拟交易平台模拟真实的期货交易市场,从微观角度观察和研究交易者和市场规律,现有的多代理平台则面临以下问题:第一,设计实现时间久远,代理的类型不易扩展。如1999年左右提出的SWARM平台等,并不能满足当前金融研究的需求,代理不够智能化,在平台运行过程中不能添加新的代理类型,代理种类单一。第二,这些平台多为单机程序,导致模拟的代理数量有限,实验结果不具有代表性。第三,实验结果数据不够详细,没有记录交易代理的交易行为。当前平台大多只存储形式简单的log数据等,未储存能够支持研究者进行详细的分析和挖掘的详细交易记录。本文针对现有平台的不足,对满足当前需求的多代理交易平台中用户交易行为分析的支撑技术进行了研究,主要完成了以下研究内容:首先,本文提出支持交易代理(Agent)类型扩展的设计方案。本文将交易代理分为代理端和交易策略两个模块,代理端只需启动交易代理的生命周期等,而交易代理交易策略的多样性则由交易策略模块实现。系统中不同类型的代理将模拟不同的真实交易者。用户可以通过类似spring注入方式添加新的交易代理类型和交易策略。其次,本文设计实现平台采用C/S架构,灵活增加节点,解决代理数量瓶颈问题。分布式的多代理平台可以模拟更多的交易者,带宽等都影响代理提交订单先后顺序,更加真实且符合实际。同时,本文利用Java NIO解决大量交易代理与Server端实时交互问题。Agent端使用Non-blocking IO与Server通讯,在这个阶段不涉及数据库的操作,布设在同一个节点上的Agent与Server的通讯可以并行,进而缩短了交易时间。最后,本文提出了混合存储架构。通过将NoSQL技术融入到存储架构中与关系型数据库一起组成混合的持久层框架,来存储海量的期货交易记录。混合存储架构十分灵活,可以在添加不同的持久性存储解决方案的同时不改变客户端软件。最后海量的历史代理交易行为log和市场交易信息可以支持进一步的分析和挖掘,从而揭示金融市场规律和交易者在不同环境下的交易特征。为了验证本文研究的合理性,本文设计实验分别对基于Java Non-Blocking IO的方案的效率和RDBMS与NoSQL混合存储架构的可用性及效率进行了验证。本文通过模拟不同的测试场景来对比Blocking IO和Non-Blocking IO能够支持的交易代理数量和请求响应时间。实验证明通过Non-Blocking IO能够支持更多的交易代理交易请求和更快的响应速度。本文还通过BenchMark对存储在不同数据库和存储架构中的数据进行压力测试,对比了他们之间读写操作响应时间等,验证了RDBMS与NoSQL存储架构高可用性和存储效率。
[Abstract]:For the complicated financial market, such as the Chinese stock index and futures market, the market price fluctuation and the relationship between investors, such as speculative trading, are particularly important to understand the law of the market transaction and the design of the policy. But the influence of the traders with different trading characteristics and their trading patterns is still unclear. We can simulate the real futures trading market based on the Agent-Based Modeling design and implement the simulation trading platform. We can observe and study the traders and the market rules from the micro point of view. The existing multi-agent platform is faced with the following problems: first, the design is long and the type of agent is not easy to expand. For example, about 1999 The SWARM platform and so on can not meet the needs of the current financial research, the agent is not intelligent enough, the new agent type can not be added to the platform operation, and the agent type is single. Second, these platforms are mostly single computer programs, which result in the limited number of simulated agents and the experimental results are not representative. Third, the experimental results are not detailed enough. The current platform mostly only stores simple form of log data, and does not store detailed transaction records that can support the detailed analysis and mining of the researchers. This paper, aiming at the shortage of existing platforms, supports the support technology of user transaction behavior analysis in the multi-agent trading platform that meets the current requirements. The main contents of the study are as follows: first, this paper proposes a design scheme to support the type extension of the transaction agent (Agent). In this paper, the transaction agent is divided into two modules: the proxy and the transaction strategy. The agent only needs to start the life cycle of the transaction agent, while the diversity of the transaction agent transaction strategy is based on the transaction strategy. Module implementation. Different types of agents in the system will simulate different real traders. Users can add new transaction agent type and transaction strategy by similar spring injection. Secondly, the design and implementation platform uses C/S architecture to flexibly increase nodes and solve the bottleneck problem of the number of agents. The distributed multi-agent platform can be modeled. More traders, bandwidth and so on affect the order of proxy submission order, more real and practical. At the same time, this paper uses Java NIO to solve the real time interaction problem between large transaction agents and Server side, and uses Non-blocking IO to communicate with Server in the.Agent terminal, which is not involved in the operation of the database at this stage and is laid on the same node. The communication between Agent and Server can be parallel and thus shorten the transaction time. Finally, this paper proposes a hybrid storage architecture. By incorporating NoSQL technology into a storage architecture, a hybrid persistent layer framework is formed with relational databases to store a large number of futures trading records. The hybrid storage architecture is flexible and can be added to different types of storage architecture. The persistent storage solution does not change the client software at the same time. Finally, the massive historical agent trading behavior log and the market transaction information can support further analysis and mining, thus revealing the rules of the financial market and the trading characteristics of the traders in different environments. Do not verify the efficiency of the Java based Non-Blocking IO scheme and the availability and efficiency of the RDBMS and NoSQL hybrid storage architecture. This article compares the number of transaction agents and request response times that the Blocking IO and Non-Blocking IO can support by simulating different test scenarios. The experiment proves that the Non-Blocking IO can be supported by the Non-Blocking IO. More transaction agent transaction requests and faster response speed. This paper also tests the data stored in different databases and storage architectures by BenchMark, compares the response time between read and write operations, and verifies the high availability and storage efficiency of RDBMS and NoSQL storage architecture.

【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP311.52;TP391.9

【相似文献】