本文从模型拟合角度出发，选择了非线性结构的变系数模型对股价进行估计。并对比FF三因子模型、计量模型，对我国沪市A股市场的超额收益进行对比讨论，研究模型解读动量与反转效应的能力。我将考察期分为短期（1个月、2个月），中期（3个月、6个月），长期（1年、2年）。 第一部分我利用对数市值纳入变系数部分，使用模型拟合结果对市场筛选后共288支股票进行检测。我发现，市场普遍存在反转效应。简单的模型结果解释出现一定程度失灵。 第二部分我对个股进行特征研究。首先分散选取12支股票，使用时间作为变系数解释变量，发现部分存在显著的变系数特征。绘制预测结果发现通过显著性的变系数模型在剧烈的波动处有明显改良。使用动量策略发现，模型估计更平滑，对股价运动趋势估计与事实较为接近。平均来看，形成期为短期1个月与长期1年，对超额收益的预测能力较好，随着形成期变动，模型与真实出现背离。 第三部分我对12支股票逐一优化，选择最优的变系数结构与线性结构。加入了个股因素变量、公司变量、投资者变量。结果指出，很多变量对风险系数beta具有影响。举例发现，变系数模型在部分时间点会对估计误差有修正的作用。估计结果是有偏且不是全局优化。使用优化模型进行估计预测后进行效应对比，发现长期的估计能力大大提高。对未来研究的趋势特征更为接近有效。 最后我对比了全局使用模型结果与结合历史数据的策略差异。我发现，使用真实历史数据可以得到更优的估计水平。这一点也很好的指向了一个论述。当市场信息还未完全开发，简化假设下的模型不能准确描述市场行为。 从研究中得到的启示建议。第一，优化模型可以更加接近真实序列趋势，但是外生的信息会使得真实特征（本文研究的动量与反转效应）无法被模型解释。第二，在实际股价预测中，变系数模型由于其连续性的特点，具有一定的实用价值，计量模型倾向于高估过去的赢家，变系数模型可以得到更合理的结果。

In this paper, from the perspective of model fitting, the varying coefficient model of nonlinear structure is selected to estimate the stock price. And comparing with the FF three-factor model and the econometric model, we discuss the excess returns in Shanghai stock market and study the model's ability to interpret the momentum and reversal effects. We divide the study period into short-term (1 month, 2 months), medium-term (3 months, 6 months), long-term (1 year, 2 years). In the first part, we fit the logarithmic market value into the model. Use results of model fitting to test 288 stocks after market screening. We have found that there is a widespread reversal effect in the market. The simple interpretation of the model results in a certain degree of failure. In the second part, we study the characteristics of individual stocks. Firstly, 12 stocks were randomly selected. Using time as an explanatory variable of varying coefficients, it was found that there were significant variable coefficient characteristics. Drawing the prediction results shows that significant changes in the coefficient of variation have significantly improved the point of fierce fluctuations. Using a momentum strategy, model estimates are smoother and the estimates of stock price movement trends are closer to the facts. On average, 1-month formation period or 1-year formation period predicts the excess returns better. With the increasing of the formation period, the models deviate from the real situation. In the third part, we optimize each of the 12 stocks, and choose an optimal varying coefficient structure and a linear structure. We add stocks factor variables, company variables, and investor variables. The results indicate that many variables have an impact on the risk factor beta. For example, it is found that the varying coefficient model has a correction effect on the estimation error at some points. The estimation result is biased and not globally optimized. After comparing the effects of the estimation model using the optimization model, it is found that the long-term estimation ability is greatly improved. The trend characteristics of future research are more effective. Finally, we compared the strategy differences between use of model results and historical data. We have found that using real historical data can lead to better estimates. This is also a good phenomenon to a discussion. When the market information has not yet been fully developed, the model under simplified assumptions cannot describe the market behavior well. Revelation suggestions were obtained from the research. First, the optimization model can be closer to the real feature trend, but the exogenous information will make the real features (momentum and reversal effect studied in this paper) unable to be interpreted by models. Second, in actual stock price forecasting, a varying coefficient model has certain practical value because of its continuity characteristics. The measurement model tends to overestimate the winners in the past, and the varying coefficient model can get more reasonable results.