中国股市的Fama-French五因素模型实证研究：噪声定价与盈利能力

需积分: 10 74 浏览量更新于2024-07-15 1 收藏 410KB PDF 举报

本文主要探讨了"Fundamental_Factor_Models_Using_Machine_Learning.pdf"文件中关于Fama和French提出的五因素资产定价模型在中国市场的应用和实证评估。三位来自Robeco的专家对该研究表示了高度关注，因为Fama和French在过去的文献中对资产定价理论做出了重大贡献。然而，他们也对新近的研究持批判态度，认为需要深入分析其有效性。 Fama和French在2015年的论文中提出了一种扩展的五因素模型，该模型在传统三因素模型（规模、价值和动量）的基础上，加入了与盈利能力相关的新因素以及与投资风格相关的新变量，旨在纠正价格噪声带来的平均回报偏差。研究者对中国股市从1997年至2015年的大量数据进行了详尽的分析，结果显示，五因素模型在中国市场中的表现持续优于传统的三因素模型。然而，与Fama和French的研究结果有所不同的是，中国市场的实证分析发现，价值和盈利能力这两个因素对于解释平均回报至关重要，而投资风格因子在描述样本内的平均收益时显得冗余。这表明在中国特定环境下，传统的投资策略可能需要更多地考虑公司的基本面，而非仅仅依赖于传统的市场宽基指标。此外，对于按照价值和增长双排序的左半边（LHS）投资组合，五因素模型的一个主要问题是未能完全捕捉到高平均回报的特性。这提示了在构建投资组合和策略时，除了考虑传统的风险调整因子外，还需要对市场噪声和公司基本面的深入理解。这篇论文揭示了在中国市场背景下，五因素模型相较于三因素模型具有一定的优势，但也指出了模型在某些方面的局限性。这对于投资者和研究人员来说，提供了一个关于如何在中国资产定价中结合机器学习和更深层次的基本面因素进行优化决策的重要参考。同时，这也预示着未来资产定价模型的发展可能需要进一步探索噪声处理和盈利能力等新的考量维度。

price from underlying value, imparts bias to estimates in mean returns and return premia. To mitigate the effects of noisy prices

on empirical regressions, we follow their suggestion for bias correction by constructing the VW portfolios, using the prior month

market cap as the weighting variable. The six portfolios that are double sorted based on Size and B/M are denoted as SH, SN, SL,

BH, BN, and BL, and the value factor HML is calculated as [(SH − SL) / 2 + (BH − BL) / 2]. The proﬁtability and investment factors

of the 2 × 3 sorts, RMW and CMA, are deﬁned similarly using OP and Inv instead of B/M as the second sorting criterion. The Size

factor, SMB, is the average of the three stock portfolio returns (SMB

B/M

,SMB

,andSMB

Inv

), which are constructed by averaging

the difference between three small and large B/M, OP, and Inv stock portfolio returns for each month, respectively. As a proxy for

the excess market portfolio return, Mkt, we use VW monthly return including all A-share stocks listed on SHSE and SZSE in excess

of the monthly deposit interest rate calculated using the one-year deposit interest rate.

The second approach in our tests is for portfolios constructed from the 2 × 2 sorts on Size and B/M, OP, or Inv, where the

breakpoint for MBM and GEM A-share stocks is the median MBM market cap and the other three ﬁrm characteristics. Similar

to those factors from the 2 × 3 sorts

, factors from the 2 × 2 sorts are also roughly neutral with respect to Size. However, the

HML, RMW and CMA factors are only neutral with respect to their own ﬁrm characteristics, as they are constructed without con-

trolling for the other two ﬁrm variables. For example, the average HML return is a mix of premiums related to B/M, OP, and Inv

since it is constructed without isolating the effects of the latter two variables. To further jointly neutralize the effects of four ﬁrm

characteristics, the return on factors is constructed from the 2 × 2 × 2 × 2 sorts on Size, B/M, OP, and Inv. We construct 16 VW

portfolios from the intersections of two Size groups, two B/M groups, two OP groups, and two Inv groups using MBM medians as

breakpoints. Then, factors from the 2 × 2 × 2 × 2 sorts are calculated as the differences between average returns on eight port-

folios (small minus big, high minus low B/M, robust minus weak OP or conservative minus aggressive Inv). Unlike factors from

the 2 × 3 or 2 × 2 sorts, factors from the 2 × 2 × 2 × 2 sorts are roughly neutral with respect to the other three variables.

The SMB factor, for example, is roughly neutral with respect to value, proﬁtability, and investment. In other words, it better iso-

lates the premiums in average returns related to B/M, OP, and Inv. One should note, however, that the ﬁve-factor exposures for

HML, RMW, and CMA from the 2 × 3 and 2 × 2 sorts may isolate the premia in average returns related to value, proﬁtability, and

investment as effectively as the factors from the 2 × 2 × 2 × 2 sorts, as multivariate regression coefﬁcients estimate marginal ef-

fects. Moreover, there is no reason to expect that neutrality with respect to ﬁrm characteristics may lead to low correlation be-

tween factor returns. Details on the construction of factors is provided in Table A.1.

2.2.2. LHS portfolios

We construct 25 Size-B/M, Size-OP, and Size-Inv portfolios to use as LHS assets in asset pricing regressions. The Size, B/M, OP,

and Inv quintile breakpoints use only MBM stocks, while the sample includes all MBM and GEM stocks. The 25 VW double-sorted

portfolios are then calculated using prior month VW monthly excess returns from July of year t to June of year t +1,rebalancing

the portfolios at the end of June each year. Similar to the problem of factor construction shown above, the Size-B/M, Size-OP, and

Size-Inv portfolios do not isolate value, proﬁtability, and investment effects in average returns because their ﬁrm characteristics

(B/M, OP, and Inv) are correlated. Following FF (2015a), we construct portfolios with 2 × 4 × 4 sorts on Size and pairs of the

other three variables (B/M, OP, and Inv). To spread stocks more evenly, the portfolios are allocated to two Size groups (Small

and Big) using the MBM median market cap breakpoint. Stocks in each Size group are then allocated independently to four B/

M, OP, or Inv groups using MBM quartile breakpoints speciﬁc to the Size group.

2.3. Summary statistics for factor returns

Panel A of Table 1 shows descriptive statistics for the means, standard deviations and t-statistics of monthly factor returns. In U.S.

data for 1963–2013, FF (2015a) ﬁnd that the factors for three versions appear with signiﬁcantly positive risk premiums. In our sample,

however, we ﬁnd that only the mean monthly time series premium for the SMB factor is signiﬁcantly different from zero. Size pre-

miums (average SMB returns) for China for 1997–2015 are large, 1.014% per month (from the 2 × 3 sorts) to 1.041% per month

(from the 2 × 2 sorts) for the three versions of the factors. Recall that each factor except Mkt from the 2 × 2 and 2 × 3 sorts controls

for Size and one other ﬁrm characteristic. The factors from the 2 × 2 × 2 × 2 sorts control for all four ﬁrm characteristics. Joint controls

have little effect on SMB. The standard deviations of SMB are similar, but with more than 3.6 standard errors from zero, larger than

those in FF (2015a) (more than 2.3 standard errors from zero). The correlations of the different versions of SMB are also very high,

from 0.986 to 1.000, largely because the Size breakpoint for the three versions of SMB is the MBM median market cap, and they

use all stocks for factor construction.

The three versions of the Mkt factor have large positive equity premiums (average Mkt returns) (insigniﬁcant), 0.833% per month

(t = 1.440). The results in Table 1 show that the descriptive statistics for the HML, RMW, and CMA factors rely on the version of factor

portfolio constructions. The average HML, RMW, and CMA returns are larger in the 2 × 3 sorts than the 2 × 2 and the 2 × 2 × 2 × 2 sorts

mainly because in former versions, stocks in the middle 40% of B/M, OP, and Inv are excluded for factor construction. For example,

the average HML return is 0.132% per month in the 2 × 3 sorts versus 0.060% per month in the 2 × 2 sorts and 0.041% per month in

the2×2×2×2sorts.Proﬁtability premiums (average RMW returns) for China are low, 0.021% per month in the 2 × 3 sorts, 0.011%

per month in the 2 × 2 sorts, and even become slightly negative in the 2 × 2 × 2 × 2 sorts (− 0.009% per month). Investment premiums

Gray and Johnson (2011) construct the AGfactor (a factor-mimicking portfolio based on asset growth) that is neutral with respect to B/M rather than Size. In an

untabulated analysis, we replicate their approach by constructing the proﬁtability and investment fact ors neutral to B/M, along the lines of Gray and Johnson

(2011). The results remain robust and are available upon request.

145Q. Lin / Emerging Markets Review 31 (2017) 141–163

剩余22页未读，继续阅读

Quant0xff

粉丝: 1w+
资源: 459

中国股市的Fama-French五因素模型实证研究：噪声定价与盈利能力

Barath_Five-Point_Fundamental_Matrix_CVPR_2018_paper.pdf

Machine Learning - Tom Mitchell

写一个使用基本面分析和因子分析来筛选出具有潜在投资价值的个股的代码

用Python求取语音的基音频率图

用Python求取语音的基音频率

python 基本面选股

fft求谐波含量matlab

用Python画基音的基音频率图

matlab fft 後如何取出基波波形

最新资源