实物交易指数的深度学习：经济定价、统计识别和数据驱动

发布时间：2017-12-27 09:34

本文关键词：实物交易指数的深度学习：经济定价、统计识别和数据驱动　出处：《浙江工商大学》2017年博士论文　论文类型：学位论文

【摘要】：中国轻纺城市场是中国纺织品实物交易的重要市场.作为纺织品实物交易风向标的柯桥纺织指数,发布于2007年10月,至今已有九年多,其影响深刻而广泛,一直是行业和社会关注的焦点.对价格指数进行定量分析,探究纺织指数本身,以及纺织行业的发展,有着十分重要的意义.但是,至今仍未见有系统的定量分析和研究.本文以技术面为导向,利用纯技术指标,基于大数据分析理念,把描述性统计过渡到推断性统计,把经典的情景分析导入可检验的模拟预测中.从经济定价到传统的统计识别,直至数据学习,重构纺织价格指数.本文的研究从经济定价到传统的统计识别,即从局限的经济定价要素到苛刻的统计假设条件,然后是完全放宽的数据驱动模式,最后是数据学习,到实时预测,这样一条主线展开.以计量经济学、统计学和数学多学科交叉的方法研究纺织价格指数.利用经典的资本资产定价模型对纺织价格指数做经济意义上的定价.用统计检验方法推断纺织价格指数总体的性质.据此,利用完全数据驱动的非参数回归的性质,对影响纺织指数的因素进行筛选和识别,在此基础上建立半参数时变系数回归模型,分析各因素在整个指数中所占的比重.再用经典的简单模型和时间序列模型对纺织价格指数做统计学意义上的定价或预测.最后建立状态空间模型,利用卡尔曼滤波实时预测纺织价格指数.对比分析法和描述性研究法贯穿整篇文章的实证分析.本文建立不同的模型对纺织价格指数进行模拟,通过描述性研究,把模拟的结果展示出来;通过对比分析,说明各种因素对纺织指数的影响程度,以及不同模型用于纺织指数的预测时的优劣.用探索性研究法确立总体模型的参数,通过非参数路径分析法揭示非线性变量的变动对被解释变量的直接的和间接的影响.本文的主要研究内容如下:(1)梳理金融资产、实物资产的定价模型,特别是资本资产定价模型(CAPM)的应用及其模型假设的检验.(2)用经典的资本资产定价模型对纺织价格指数进行经济学意义上的定价.由于现实中没有符合CAPM模型假设中的无风险收益,所以用方差最小的与市场组合零协方差证券组合的期望收益作为无风险收益建立零贝塔CAPM模型,检验不同类型的产品市场是否适用CAPM模型,进一步地测算不同类型产品的市场风险和检验其定价的合理性.(3)对纺织价格指数时间序列的样本进行统计分析和检验.采用描述性统计分析,正态性检验,平稳性检验,序列自相关检验,游程检验,独立同分布检验.所有的数学模型,包括统计模型都有其适用的条件和范围,在对实际问题建立模型之前,必须通过样本检验模型假设是否成立,即总体是否具有这些性质,否则所建模型是不可靠的,得到的结论也没有实际价值.(4)用数据驱动的方法建模.应该没有完全符合模型假设的实际问题,常常只是近似满足,近似程度越高,模型的可靠性越好,所以本文采用条件相对宽松的完全数据驱动的非参数方法.首先基于大数据的理念,确定可能直接影响纺织价格指数的因素集,从中筛选影响显著的因素,并对这些具有显著影响的因素进行识别,区分其对纺织价格指数的影响是线性的还是非线性的,然后建模分析各因素对纺织指数影响的强度.(5)建立预测模型.一个合理的模型体系应该具备预测功能.CAPM模型有严苛的难以检验的假设和现实中很难满足的前提,而用半参数模型进行预测之前,必须先预测线性的和非线性的自变量的值,所以都只有理论上的预测功能.最后建立状态空间模型,利用Kalman滤波,实现对纺织价格指数的实时预测.我们取2007年5月到2016年12月作为样本期间,通过研究得到如下结论:(1)经济定价的结果.我们把CAPM模型用于实物交易指数的定价,分别对纺织价格的周指数和月指数建立CAPM模型,发现对月指数建立的模型好于对周指数建立的模型.在统计意义上,后者方程的决定系数较大;在经济意义上,方程的常数项不能拒绝为0的原假设,即认为定价合理.一次项系数显著不为0,即价格受市场风险的影响,其中的超额收益是市场风险溢价.但是对30个中类指数建模的结果发现,只有9个中类月指数模型的常数项不能拒绝为0的原假设,且常数项系数显著大于0.考虑到现实中并不存在可以无限借贷的无风险利率,用零贝塔CAPM模型再对各中类市场进行检验,计算得5个大类市场,即原料、坯布、服装面料、家纺和服饰辅料市场在样本期内的零协方差证券组合的月期望收益率的估计值分别为0.9479%、0.1448%、0.1037%、-0.0483%、0.1535%,样本期内一年期存款基准利率的月化利率是0.125%,坯布市场、服装面料市场和服饰辅料市场的零协方差组合的月期望收益率与之相当,而家纺类的市场零协方差组合的月收益率小于零.零贝塔CAPM模型的估计结果与CAPM模型的估计结果相差不多.究其原因,应该是中国轻纺城纺织品交易市场以及市场经营者对市场的交易的期望与CAPM的假设条件相去甚远,所以需要通过样本揭示纺织价格指数总体的性质.(2)统计识别的结果.基本统计分析的结果认为纺织价格指数厚尾右偏,不服从正态分布.ADF检验和KPSS检验的结果认为纺织价格指数序列存在一阶单位根;自相关性检验的结果认为存在序列自相关性;游程检验的结果认为序列不是随机游走过程;BDS检验的结果确实不是独立同分布的.所以,纺织价格指数的性质并不适合前提严格的模型.(3)数据驱动的变量选择和识别的结果.首先依据产业经济学和宏观经济学理论,选取了 14个可能影响纺织品价格指数的变量,以 LCLS(Local Constant Least-squared)方法估计方程,并用 LSCV(Least-Squared Cross-Validation)法选取各控制变量的窗宽,认为窗宽大于两倍样本标准差的6个控制变量对因变量没有显著影响,将其剔除.对余下的控制变量做识别,以LLLS(Local Linear Least-Squared)方法重新估计方程,仍用LSCV法选取各控制变量的窗宽,认为其中窗宽大于两倍样本标准差的4个控制变量对因变量的影响是线性的,其余变量对因变量的影响是非线性的.筛选的结果是,原油平均现价、商品零售价格指数、工业品出厂价格指数、因变量的滞后一阶变量、棉花“指数A”、货币供应量、金融机构企业存款、金融机构短期贷款、汇率水平和期末国家外汇储备等变量对纺织品价格指数有显著影响;识别的结果是,原油平均现价、商品零售价格指数、工业品出厂价格指数、以及因变量的滞后一阶变量4个变量的窗宽大于两倍的样本标准差,认为这4个变量对纺织品价格指数的影响是线性的.棉花“指数A”、货币供应量、金融机构企业存款、金融机构短期贷款、汇率水平和期末国家外汇储备对纺织价格指数的影响是非线性的,称之为控制变量.根据选择和识别的结果,建立半参数变系数回归方程,用来模拟各非参数变量对被解释变量的影响强度.(4)半参数模型的模拟结果.建立关于线性变量和非线性变量的半参数回归方程,并且计算各个参数项和包括7个控制变量的非参数项在纺织价格指数中所占的比重.4个参数项在纺织品价格指数中所占的比例近80%,非参数项则占20%强.在所有6个控制变量中,棉花“指数A”起着主要的作用,实证和模拟的结果认为,在样本期内如果棉花“指数A”不超过两倍的样本均值,可以使得各部分的占比不出现极端值,不超过平均占比的两倍.(5)建立实时预测模型.用三类模型,即简单模型、时间序列模型和状态空间模型(Kalman滤波)对纺织品价格总指数进行样本期内预测(估计)和一期外推预测,在样本期内,用均方根误差衡量预测误差,误差均在可接受的范围内,而时间序列模型的预测误差比简单模型要大,说明时间序列模型的设定正确.状态空间模型的递推预测结果受初始值的影响,但是当递推次数足够多以后,初值的影响会消除,所以舍弃状态空间模型预测序列的最初60期以后,预测精确度高于时间序列模型.无论从预测精确度还是可计算性,这三个模型中状态空间模型是最优的。
[Abstract]:Chinese Textile City market is an important market for textile Chinese barter barter. As the textile barometer of Keqiao textile index, released in October 2007, has been more than nine years, its profound and extensive influence, has been a focus of concern in the industry and society. The price index for quantitative analysis, research and development of textile and textile index itself. The industry, has a very important significance. However, quantitative analysis and systematic study has yet to surface. In this paper, technology oriented, the use of pure technical indicators, based on the concept of big data analysis, the descriptive statistics to inferential statistical analysis, the classic scenario simulation test. The prediction of import from to identify the traditional economic pricing statistics, until the data reconstruction study, textile price index. This paper from the economic pricing to traditional statistical recognition, namely from the Bureau of economic limit set The price of the elements to the statistical hypothesis of harsh conditions, and then the data is completely relaxed driving mode, and data learning, to real-time prediction, such a masterstroke. In econometrics, statistics and mathematical methods of interdisciplinary research in textile price index. The economic sense of textile pricing price index by capital asset pricing the classic model. The statistical test method is used to infer the overall price index of textile properties. Accordingly, the use of fully data-driven non parametric regression, screening and identification of factors affecting the textile index, based on semi parametric variable coefficient regression model, analysis of various factors in the whole index in proportion. Simple model and time series model, and then do the classic statistical significance on pricing or prediction of textile price index. Finally, the state space model is established, using the card The Kalman filter real-time prediction of textile price index. Empirical analysis comparative analysis and descriptive research method throughout the article. This paper establish different models of textile price index is simulated by descriptive study, show the simulation results; through the comparative analysis, explain the influence degree of various factors on the textile index, and different model for forecasting the quality of the textile index. Through exploratory research to establish the overall model parameter method, reveal the nonlinear variables by non parametric path analysis changes on the explanatory variables of the direct and indirect influence. The main contents of this paper are as follows: (1) pricing model combing financial assets, real assets, especially the capital the capital asset pricing model (CAPM) test and application of model assumptions. (2) the economics of textile price index with the classical capital asset pricing model The meaning of pricing. Because the real risk-free income does not meet the CAPM model hypothesis, so with the minimum variance of the market portfolio and zero covariance portfolio expected return as a risk return model of zero beta CAPM, inspection of different types of products the market is suitable for CAPM model, further to estimate the different types of products in the market the risk and test the pricing rationality. (3) the textile price index time series samples were analyzed and tested. By using descriptive statistical analysis, normality test, stability test, serial autocorrelation test, runs test, independent and identically distributed test. All mathematical models, including statistical models have their the applicable conditions and scope, before establishing the model for the actual problem, must through the sample test model hypotheses, namely whether these properties, or model Is not reliable, the conclusion is of no practical value. (4) modeling method using a data driver. There should be no problem in line with the actual model hypothesis completely, often only approximate, approximation degree is higher, the better reliability model, so this paper uses the non parametric method of complete data conditions relatively loose driving. Firstly, based on the concept of data, determine the factors that may directly affect the textile price index set, select significant factors from, and identification of the significant influencing factors, distinguish the influence on textile price index is linear or nonlinear, then the modeling and analysis of influence of various factors on the strength of the textile index. (5) prediction model is established. A reasonable model system should have prediction function of.CAPM model is difficult to meet the premise hypothesis and the harsh reality is difficult to test the use of semi parametric Before the number of models must forecast the linear and nonlinear variable value, so only the theoretical prediction function. Finally, the state space model is established, using Kalman filter, to achieve real-time prediction of textile price index from May 2007 to December 2016. We take as the sample period, the results are as follows: (1) the economic valuation results. We put the CAPM model for real exchange index pricing, the price index of textile week and month index to establish the CAPM model, found on index models in index models built for the week. In the statistical sense, the latter decision coefficient equation is relatively large; economically, constant cannot refuse to the original equation hypothesis 0, that is reasonably priced. A coefficient is not significantly affected by the market price of 0, the risk of excess income, which is the market risk But for the premium. 30 class index modeling results show that the constant term is not only the 9 class index model to 0 original hypothesis, and the constant coefficient is significantly higher than the risk-free interest rate of 0. taking into account the reality does not exist unlimited borrowing, with zero beta CAPM model of each class in the market test, calculation of 5 big markets, which estimated the expected rate of return of the raw materials, fabric, apparel fabrics, home textiles and apparel accessories market in the sample period of the zero covariance portfolio values were 0.9479%, 0.1448%, 0.1037%, -0.0483%, 0.1535%, the sample period
【学位授予单位】：浙江工商大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：F224;F426.81;F768.1

【相似文献】