资本市场对经济发展发挥越来越重要的作用，机构投资者更是股票市场发展壮大的重要推动力量。机构投资者的大额交易、理性投资也成了证券市场交易的重要组成部分，如何能更好地提高交易效率和减少交易成本，对机构投资者而言是十分重要的问题。在交易中，对冲基金、养老基金、共同基金以及其他机构交易者广泛使用地应用算法交易。他们将大额的交易分解为若干笔小额的交易，以便更好地管理市场冲击成本、机会成本和风险。在市场上由算法交易完成的交易额中，被动型算法交易是最常用的算法交易策略，且其中绝大部分都是通过VWAP交易策略进行的。本文针对VWAP算法交易策略的基本目标和核心思想，即依据股票市场日内成交量的比重拆分大额订单进行交易，通过研究上证指数日成交量、日内成交量高频时间序列的分布特征和统计特征，我们建立了均值-方差模型，并通过消除异方差和针对长记忆性建立ARFIMA模型，在一定程度上提高了预测日内成交量序列的准确度。在论文中，我们首先对近二十年来中国股票市场的日成交量时间序列进行了统计分析，发现序列近似服从幂律分布；之后我们选取1996年到2012年的日成交量数据，经检验是平稳时间序列，而后我们结合ARMA模型和ARCH类模型，分别消除了自相关和异方差，经验证ARMA（3,1）-EGARCH(2,1)和ARMA(3,1)-PARCH(1,2)拟合效果较好。其次，我们针对日内一分钟间隔成交量高频数据做了分布特征分析，发现其呈现对数正态分布，且曲线拟合效果较好；之后对其数据进行统计分析，包括平稳性、自相关、异方差和长记忆性的研究；而后通过对原始成交量时间序列和剔除日内周期趋势的成交量序列分别建立合适的均值-条件方差模型，经检验发现剔除周期趋势很有必要，且对剔除周期趋势后的序列建立的ARMA-EGARCH（2，2）模型拟合和预测效果较好。最后，经R/S检验，我们发现日内高频成交量序列存在较强的长记忆性，分别对三个月日内五分钟间隔的原始成交量时间序列和剔除周期趋势的成交量时间序列建立ARFIMA模型，拟合和预测的效果均较好，结果表明日内成交量的长记忆性对建模有较大的影响，且剔除周期趋势后建模预测的效果更好。

With the rapid development of China’s economy, the capital market is playing an important role day by day, especially the institutional investors. They are one of the most significant parts in China’s stock market, and their block trading and rational investment strategy have become very important power for the stable and healthy development of stock markets. For them, how to increase the efficiency of trading and reduce trading costs are vital issues. Talking about the process of trading, we need to mention algorithmic trading, which is very popular among hedge funds, pension funds, mutual funds and other institutional investors. They divide large block of trading into small pieces, and trade in the market within periods of time to reduce market impact costs, opportunity costs and risks. Of all the trading in stock markets, passive algorithmic trading is most commonly used, and most of the institutional investors use VWAP strategy.The core idea of VWAP strategy is to trade according to the distribution rate of intraday volume series. By examining the statistic characteristics of Shanghai Security Index daily volume time series and intraday volume time series, we set up mean-variance model. By eliminating autocorrelation, heteroscedasticity and long memory feature, we demonstrate that the simulation and forecast performance is improved to some extent.In this paper, firstly we study the statistic feature of daily volume series and find out that they fit power law distribution, and then we set up mean-variance models to simulate, which indicates that ARMA（3,1）-EGARCH(2,1) and ARMA(3,1)-PARCH(1,2) have the best simulation performance.Secondly, we put forward a modeling method for the intraday volume series. By comparing the simulation and forecast outcome of modeling of the original volume data and those which have been removed intraday periodic volume components, we demonstrate that the second method has better outcome in stimulation and forecast for the intraday volume time series, and ARMA-EGARCH (2, 2) outperform others in many aspects.Finally, by R/S test, we find out that there is long memory feature in the intraday high-frequency volume time series. We set up ARFIMA model to simulate and forecast the original data and the data which has been removed the W pattern. The results indicate that the long memory feature of intraday volume series has great influence on the performance of simulation, and the forecast results with periodical pattern removed turns out better outcome.