Research On Order Strategy And Investment Portfolio Based On Market Microstructure And Machine Learning Algorith

The globalization of financial competition is becoming increasingly fierce,the prosperity of the financial industry has significant contributions to social progress and stability.The research of inherent development laws and characteristics under the microstructure of the financial market from a scientific perspective can provide investors and decision makers with a scientific basis for investment decisions,and promote the prosperity,healthy and orderly development of the financial market.We conduct research on the microstructure theory,application,portfolio solution and optimization of financial markets,and investor activities,hoping to provide some valuable contributions through a series of quantitative models and quantitative analysis methods.The rapid development of electronic information technology and computing science has transformed the trading mechanism of the global financial market from a quote-driven system to an order-driven electronic trading system.At present,most stock and securities futures and other financial derivatives trading sites in the world use quote-driven electronic trading systems.In Chinese financial market,such as Shanghai,Shenzhen Stock Exchange and various futures exchanges also use this system.The theoretical and empirical research on the microstructure of the micro market has also changed from a quote-driven market trading system to an order-driven market trading system.A complete,concise,and effective quantitative model can clearly describe the various phenomena and price phenomena in the order-driven market microstructure,and guide investment and regulatory activities.Therefore,it has attracted extensive attention from researchers.Based on the market microstructure,we mainly studied the following issues(Chapters 2-4): continuous time stochastic dynamic model,the impact of order flow imbalance on price,and the optimal pending order strategy based on the Monte Carlo method.Chapter 2 refers to the research on the order book model,from the initial simple static model and basic statistical feature description,to the dynamic model to describe the form,characteristics,key attributes and update process of the order book.Chapter 3 of the article introduces the continuous-time dynamic stochastic model proposed by Cont et al(2010),and uses historical transaction data on the Chinese stock market as sample data to calculate the parameters involved in the model.In Chapter 3,we study the price impact of order events,liquidity,and imbalance of order flow in order-driven market.Price volatility is one of the important issues which have long-term concerned by researchers and investors.Early studies have done a lot of work about the impact of trading imbalances on prices.Cont(2011)proposed a linear model of the relationship between order flow imbalance and price fluctuations.Considering that there is a certain difference between the microstructure of the Chinese market and the mature markets of Europe and America,we conduct an empirical analysis of the relationship between price volatility and order flow imbalance by historical transaction data in the Chinese stock market.We try to explain the elationship between the price volatility and the order flows imbalance of China Financial Market.We find that this model does not fit very well as we expected.We believe that this may be related to the high noise of high-frequency trading data,the time series nature of financial data,and the immature development of China’s stock market.In addition,the high-frequency trading data in the US financial market is a real-time snapshot,while the three-second snapshot is taken in the Chinese stock market.There are differences in the data structure,which may also be the reason why the model interpretation effect is not ideal.Subsequent research can also be considered from the perspective of the non-linear correlation between order flow imbalance and price volatility.In Chapter 4,we study the Optimal order placement strategy use the Monte Carlo method.Order placement strategy is one of the important issues in LOB research.Strategy research is a problem of highly overlapping market microstructure theory research and investment practice,which is also a concern in our research.Previous researchers have also done a lot of related work on this issue.Cont(2010)propose the Laplace transform to calculate the probability of the next order time,but they do not carry the empirical research.We propose two theorems based on the distribution of order event arrival time:(i)an exponential random variable event follows the exponential distribution when the next order event arrives;(ii)the order event type is determined according to the time of the order arrival rate.Under this framework,we simulate the birth and death process of each price level in the order book and the update process of the order book according to the definition and characteristics of the six types of orders.We describe the order event as a multi-queue service system subject to the Poisson process for modeling.The model describes the dynamic evolution of the order book by simulating the birth and death process of each price level.In this simulation process,it can effectively capture every update of the best quote.We set purchase an order for a unit in the program.We determine the effectiveness of the strategy based on the average transaction price and sample variance output by the experiment.We found that the precise simulation of the complete path of the limit order book within the specified time provides an accurate prediction of the expected transaction price of a single limit order at each position of the limit order book.The simulation algorithm provides a reference for investors to place the best order in the limit order book,the algorithm is robust,stable and accurate.In Chapter 5,we propose a commodity futures portfolio strategy based on machine learning methods.As machine learning algorithms do not require preset models and powerful learning capabilities,they are very good at predicting financial time series.We compared the performance of three classic machine learning algorithms(LR,SVM,XGBoost)in commodity futures yield prediction,and determined the machine learning model based on their performance.In addition,considering the leveraged nature of futures,we propose a futures portfolio strategy through Markowitz theory.We found that the portfolio strategy can effectively reduce the risk and maximum retracement rate,and increase the yield,Sharpe ratio.This portfolio strategy performs well.In Chapter 6,we propose several numerical acceleration methods in portfolio applications,and use the special mathematical structure of the model constraints to optimize the model.The modern portfolio solution and optimization problem is essentially a mixed integer convex quadratic programming problem.When solving the model directly,there will often be pauses caused by the slow solution speed and the solution will not continue.The previous researchers have done a lot of improvement work around the covariance matrix(input)and algorithm(output).Based on their research,we use the special mathematical structure of constraint matrix to propose three numerical acceleration methods: Method 1,Method 2,Method 3,where Method 1,Method 2 are equivalent conversions,Method 3 is not equal Price conversion.Method 1 solves the model by Cholesky decomposition of the covariance matrix in the constraints,which greatly improves the solution speed;Method 2 is divided into two steps: Method 2(Form 1)and Method 2(Form 2).In Method 2(Form 1),we perform a cone decomposition of the covariance matrix,a single cone is divided into a low rank covariance matrix and a special risk diagonal matrix.We used LDL decomposition to make the covariance matrix low dimension and reduce the storage space.In numerical tests,the speed has been increased by approximately 43 times.Method 2(Form 2)uses the characteristics of the 0-1variable in the constraint to convert the cone constraint into a linear constraint through equivalent conversion,which improves the speed by about 43%.In the numerical test,In the numerical test,the final solution model is improved by about 75 times,which is about two orders of magnitude higher than the direct solution of the initial model.In addition,considering that some investors and investment institutions are more sensitive to time(such as high-frequency traders),they lose a certain accuracy(about 10 ^-4 in numerical experiments).By adopting the method of convex relaxation proposed in Method 3 to perform non-equivalent conversion,the integer constraint is removed,but the sparseness of problem solving is controlled by increasing the handover constraint.In the experiment,we also found that it is indeed possible to maintain the sparseness of the solutions obtained.This method improves the solution speed by about 42% on the basis of Method 2.In the numerical experiment using the convex relaxation method,the average solution time is about 9.4to 9.5 seconds.In actual large-scale stock investment portfolios,especially high-frequency trading,profit opportunities often arise and disappear within a short period of time,and the speed of solution often becomes the bottleneck restricting high-frequency trading.The special mathematical properties such as sparseness of the model can significantly improve the solution performance,and our modular numerical acceleration method is also validated in the actual operation of a fund company.Chapter 7 is the trade preferences of heterogeneous traders.In the early studies on trader behavior,such as agency theory and trader behavior preference,the research was mostly based on the trader’s subjective behavior perspective.We try to study the behavior preferences of traders from the perspective of market microstructure.According to the characteristics of orders submitted by different traders,the traders are defined as fund managers driven by patience and high-frequency traders since they have status in the order book.We introduce two queue models of the order book,which describe the process of how the orders submitted by two types of traders pass through the FIFO queue.We try to explain the behavior of waiting for execution in the order queue after the order arrives in the queue.The content of this chapter is mainly a literature review of trader behavior from the perspective of micro-market structure.Subsequent empirical research will focus on this issue to further improve the research problem.Gerenally speaking,we aim to describe the characteristics of the order book in the financial market microstructure and the dynamic changes of key attributes through a series of quantitative models,in order to provide a certain reference for traders to make better trading decisions.