1. Introduction

In efficient markets, nominal prices of stocks play no role. Only market values and returns matter. In a recent paper, however, Birru and Wang (2016) show that small price stocks exhibit a peculiar behavior, when compared to large price stocks. In particular, the authors conclude that investors overestimate the room to grow for small price stocks, leading them to overestimate the skewness of future returns. Moreover, Roger et al. (2016) find that financial analysts are much more optimistic when they issue one-year ahead target prices on small price stocks, compared to their price forecasts on large price stocks. In this paper, we show that a small stock price is also associated with a high idiosyncratic volatility of the stock return.

Whichever asset pricing model used, idiosyncratic variance is the difference between total variance and variance coming from the common factors. As a result, the measure of idiosyncratic variance crucially depends on the number and the choice of common factors in the asset pricing model. In a recent paper, Harvey et al. (2015) mention that more than 300 significant factors have been uncovered in hundreds of papers, but they highlight that most claimed research findings in financial economics are likely false. Hence, we do not really know what the common factors are, what is their number, and what are their economic contents.

In this paper, we refer to the general approximate factor structures introduced by Chamberlain and Rotschild (1983) to obtain an estimate of the idiosyncratic volatility (henceforth IVOL) of stock returns. The fact that approximate factor structures take into account correlated residuals is a significant advantage of this approach. We then show that IVOL is strongly linked to the nominal price of stocks. For example, Campbell et al. (2001) use two types of common factors: the usual market factor, measured by the excess return on a broad market index like the S&P500, and industry factors which influence only subsets of stock returns. Considering only the market factor implies that residuals are correlated because of industry factors.

Our empirical study deals with approximately 8,000 stocks traded on the U.S market over a 35-year period, starting in January 1980 and ending in December 2014. When performing principal component analyses (PCA) of daily returns on the 35 non-overlapping periods of one year, we find that the average IVOL (expressed as a percentage of the average total variance) of large price stocks is lower than the corresponding IVOL of small price stocks. The result holds when we consider one factor. The reason to primarily focus our attention on the first eigenvalue comes from the numerous studies showing that a one-factor model explains an important part of the variance of returns (Trzcinka, 1986; Connor and Korajczyk, 1993; Geweke and Zhou, 1996; Jones, 2001). Our result, however, also holds when we consider three common factors.

This result is remarkable because small price stocks are not always the same over time, because of price variations and stock splits. Indeed, Weld et al. (2009) show that nominal prices have remained approximately constant in the U.S since the Great Depression. After a detailed analysis of this phenomenon through the lens of standard theories, they come to the conclusion that firms simply follow market norms, in the sense given by Akerlof (2007) in his Presidential Address to the American Economic Association, even if maintaining the price level stable comes at a cost. In particular, bid-ask spreads increase after stock splits (Copeland, 1979; Conroy et al., 1990).

Overall, we highlight that IVOL is strongly associated with the nominal price of stocks. The percentage of variance explained by common factors is much higher for large price stocks than for small price stocks. Our result is persistent over time (our analysis covers the period 1980-2014) and does not hide a small size effect because it remains valid on the tercile of large caps. Since the standard theory of finance considers that stock prices are irrelevant to describe returns, we therefore argue that the nominal stock price is a supposedly irrelevant factor (SIF), as defined by Thaler (2015). Hence, our result confirms that most small price stocks have lottery-type characteristics, which is not consistent with the efficient market hypothesis.

The paper is structured as follows. Section 2 recalls the basic results related to approximate factor structures (Chamberlain and Rotschild, 1983). Section 3 presents the data and descriptive statistics. Section 4 develops the empirical results and section 5 concludes.

2. Approximate factor structures

Approximate factor structures are an extension of the multifactor model of returns proposed by (Ross, 1976) in his seminal paper « The Arbitrage Theory of Capital Asset Pricing » (hereafter APT). This alternative factor model was first proposed by Chamberlain and Rotschild (1983) and characterized as follows. Suppose that $n$ financial securities are traded on the financial market. The returns on these securities (denoted$R_i,i=1,\text{..},n$) are square-integrable random variables defined on a probability space $\left(\Omega ,A,P\right)$. The first moments of the returns are denoted:

$E\left(R_i\right)=m_i;Var\left(R_i\right)={\sigma }_i^2;cov\left(R_i,R_j\right)={\sigma }_{ij}$
                 (1)

$V_n=\left({\sigma }_{ij},i=1,\text{...},n;j=1,\text{...},n\right)$
                     (2)

For the consistency of notations,${\sigma }_{ii}={\sigma }_i^2\text{.}$

Definition 1 The market is driven by an approximate K-factor structure if

$R_i=E\left(R_i\right)+{\sum }_{k=1}^K{\beta }_{ik}{F}_k+\widetilde{{\epsilon }_i}$
                         (3)

for$i=1,\text{..},n$where the common factors$F_k,k=1,\text{...},K$are uncorrelated zero-mean random variables with unit variance. The factors $F_k$ are uncorrelated with the residuals $\widetilde{{\epsilon }_i}$

$V_n=B_n{B}_n^{\text{'}}+D_n$
                            (4)

where the$\left(i,k\right)$element of$B_n$is ${\beta }_{ik}$, and$\left(D_n,n=1,\text{...}\right)$is a sequence of positive semi-definite matrices satisfying:

$\tilde{\lambda }=\underset{n}{\text{sup}}{g}_1\left(D_n\right)<+\infty$
                        (5)

where $g_1\left(D_n\right)$is the largest eigenvalue of $D_n$.

The variance of $\widetilde{{\epsilon }_i}$is denoted$s_i^2$. In the original APT the$\widetilde{{\epsilon }_i}$are assumed i.i.d. This assumption is relaxed in an approximate factor structure. The matrices $D_n$ need not be diagonal.

Equation (3) implies that the variance of returns of stock $i$, denoted ${\sigma }_i^2$, is equal to:

${\sigma }_i^2={\sum }_{k=1}^K{\beta }_{ik}^2+s_i^2$
                            (6)

where $s_i^2=V\left(\widetilde{{\epsilon }_i}\right)$  is the idiosyncratic variance of returns. It turns out that $s_i^2$ is the difference between the total variance of returns, and the systematic variance${\sum }_{k=1}^K{\beta }_{ik}^2\text{.}$ As a consequence, $s_i^2$ crucially depends on the number of common factors.

In empirical studies, $K$ is not known with certainty and no universal technique exists to determine $K\text{.}$ Prudence is then needed when dealing with idiosyncratic variances in such models.

Summing equation (6) over all stocks and rearranging terms leads to:

${\sum }_{i=1}^n{\sigma }_i^2={\sum }_{k=1}^K{\sum }_{i=1}^n{\beta }_{ik}^2+{\sum }_{i=1}^n{s}_i^2$
                        (7)

The left-hand side of equation (7) is the trace of the covariance matrix of returns $V_n$ which is also the sum of the eigenvalues of the covariance matrix$V_n$.

We get a CAPM-like factor model in the case$K=k=1\text{.}$ We then expect ${\sum }_{i=1}^n{\beta }_{i1}^2$ to be the first eigenvalue of $V_n$ for a sufficiently large number $n$ of stocks. The reason is simple. When a n-th stock is introduced, the covariance matrix $V_{n-1}$becomes$V_n$and the trace increases by${\sigma }_n^2={\beta }_{n,1}^2+s_n^2$. But ${\beta }_{n,1}^2$is added to the first eigenvalue due to the one-factor assumption and$s_n^2$feeds higher-order eigenvalues, including the new n-th eigenvalue. It turns out that the first eigenvalue increases without limits when$n$increases. Nothing equivalent can be expected for the higher-order eigenvalues in a one-factor model. These eigenvalues are bounded when the number of stocks tends to infinity.

The residual terms in equation (3) may be correlated in an approximate factor structure. One economic interpretation of such a structure is that common factors influence the returns of all stocks but some other factors, such as industry factors, influence only a subset of returns. In this case, the share of variance attributed to these factors is, in some sense, residual, and appears in the variables ${\tilde{\epsilon }}_i=1,\text{...},n,$. This results in non-zero correlations between these variables ${\tilde{\epsilon }}_i$s.

Campbell et al. (2001) use a two-step decomposition of returns to introduce industry factors (this two-step decomposition has also been used in studies focusing on the role of idiosyncratic volatility in Europe, see Kearney and Poti (2008). In the first step, the return of industry $j$ is written:

$R_{jt}={\beta }_j{R}_{Mt}+{\tilde{\zeta }}_{jt}$
                        (8)

where$R_{jt}$is the date-t return of industry j and$R_{Mt}$is the market return.

In the second step, the return of a given firm $i$ belonging to industry $j$ is decomposed as:

$R_{ijt}={\beta }_{ij}{R}_{jt}+{\tilde{\xi }}_{ijt}$
                        (9)

Replacing $R_{jt}$ in equation (9) by the expression in equation (8) gives:

${R_{ijt}=\beta }_{ij}\left({\beta }_j{R}_{Mt}+{\tilde{\zeta }}_{jt}\right)+{\tilde{\xi }}_{ijt}$
                        (10)

$R_{ijt}={\beta }_{ij}{\beta }_j{R}_{Mt}+{\beta }_j{\tilde{\zeta }}_{jt}+{\tilde{\xi }}_{ijt}$
                        (11)

It turns out that the residuals${\beta }_j{\tilde{\zeta }}_{jt}+{\tilde{\xi }}_{ijt}$are correlated when two firms belong to the same industry.

This example shows that approximate factor structures allowing correlated residuals are an efficient tool to test the assumption of a negative link between nominal stock prices and idiosyncratic volatility of returns. Moreover, this approach avoids the economic characterization of the common factors. Finally, Chamberlain and Rotschild (1983) provide a useful method to determine the number of common factors, summarized in the following proposition.

Proposition 2 If the market has an approximate$K$-factor structure, the$K+1$eigenvalue of $V_n$remains bounded when $n$ tends to infinity:

$\underset{n}{\text{sup}}{g}_{K+1}\left(V_n\right)<+\infty$
                        (12)

The interpretation of this proposition is intuitive. By the invariance of the trace of $V_n$, we also know that

${\sum }_{i=1}^n{\sigma }_i^2={\sum }_{i=1}^n{g}_i\left(V_n\right)$
                        (13)

When a new asset (numbered n + 1) is added, its variance${\sigma }_{n+1}^2$can be related to the preceding equation as follows:

${\sigma }_{n+1}^2+{\sum }_{i=1}^n{\sigma }_i^2=g_{n+1}\left(V_{n+1}\right)+{\sum }_{i=1}^n{g}_i\left(V_{n+1}\right)$
            (14)

Equations (13) and (14) imply

${\sigma }_{n+1}^2=g_{n+1}\left(V_{n+1}\right)+{\sum }_{i=1}^n\left(g_i\left(V_{n+1}\right)-g_i\left(V_n\right)\right)$
            (15)

We cannot expect that the idiosyncratic volatility of stock$n+1$simply feeds the last eigenvalue $g_{n+1}\left(V_{n+1}\right)$because eigenvalues are usually given in decreasing order. This reasoning generalizes the example of a one-factor CAPM-like model given before. If, however, there are $K$ common factors driving the economy, we know that the systematic risk of stock$n+1$feeds the first$K$eigenvalues thanks to proposition 2. Only the idiosyncratic variance is disseminated in eigenvalues of an order greater than$K$and the number of these eigenvalues increases when the number of stocks increases. It explains why, when the number of stocks under consideration increases, the$K$first eigenvalues increase without bounds and the eigenvalues of larger orders remain bounded.

In the empirical study of the next section, we cannot know for sure the number of common factors because the number of assets under consideration is bounded. We present the results obtained with 1 and 3 factors, which are the most likely numbers of factors over the period under consideration.

3. Data and descriptive statistics

Daily data on returns, nominal prices and market capitalizations come from the Center for Research in Security Prices (CRSP). We include in our analysis all ordinary common shares (codes 10 and 11) listed on NYSE, Amex and Nasdaq over the 1980-2014 period.

4. Idiosyncratic volatility and nominal prices

In general, the estimation of factor structures through nested principal component analyses (PCA) is performed by diagonalizing the covariance matrix of returns. Small firms, however, are more volatile than large firms and there is a positive correlation between the nominal price of a stock and the capitalization of the firm. As a consequence, using the covariance matrix to estimate the relative importance of idiosyncratic volatility among small price stocks and large price stocks would bias the results. For that reason, we chose to diagonalize correlation matrices of returns to neutralize the difference in total volatility between small and large price stocks.

Save View full size Expand inline Collapse inline

Figure 1: Evolution of the adjusted Frobenius norm of the correlation matrix of daily returns over the 35 years of the 1980-2014 period.

This figure presents the evolution of the adjusted Frobenius norm of the correlation matrix of daily returns over the 1980-2014 period.

The other advantage of this standardization is that the trace of a correlation matrix of returns is equal to the number of stocks under consideration.

4.1 The number of common factors

The number of factors of the approximate factor structure can be determined in a simple way by looking at the evolution of eigenvalues of the correlation matrix, thanks to the theorem of Chamberlain and Rotschild recalled in proposition 2. Figure 2 provides an illustration of this evolution of eigenvalues for 1999, 2008 and 2014. The horizontal axis gives the number of stocks in the PCA and the vertical axis gives the values of the eigenvalues. We represent the curves of the first three eigenvalues, ranked in decreasing order. These curves are obtained by performing nested principal component analyses, starting with 50 stocks and incrementing the number by 50 at each step until all stocks are included. The upper bold curve shows the evolution of the first eigenvalue when the number of stocks increases. The dashed (dotted) curve represents the second (third) eigenvalue.

Save View full size Expand inline Collapse inline

Figure 2: Evolution of the three first eigenvalues of the correlation matrix of returns with respect to the number of stocks in the PCA in 1999, 2008 and 2014

This figure presents the evolution of the first three eigenvalues of the correlation matrix of daily returns when the number of stocks included in the portfolio increases. Panel A corresponds to year 1999, Panel B to year 2008 and Panel C to year 2014. The upper bold curve corresponds to the first eigenvalue, the dashed curve corresponds to the second eigenvalue, and the dotted curve provides the values for the third eigenvalue. These curves are obtained by performing nested principal component analyses, starting with 50 stocks and incrementing the number by 50 at each step until all stocks are included (the curves for the 35 years are available upon request).

The three graphs correspond to very different years in terms of market returns. 1999 was the last year of the dotcom bubble, 2008 the year of Lehman Brothers’ bankruptcy, with a yearly market drop close to 50% and an average cross-sectional volatility of 77%. Finally, 2014 was a normal year, with an average return of 3% and a 33% average volatility (see Table 1). The three graphs perfectly illustrate our discussion about the evolution of eigenvalues. As a result, the most likely number of common factors is 1 in a number of periods. Harding (2007), however, showed that considering one factor is often a biased estimate of the true number of factors because of the finite sample properties of the correlation or covariance matrix. We also, therefore, perform a robustness check with three factors to cover the main cases. Since the number of stocks is much larger than the number of days, three factors is a reasonable choice (Bai and Ng, 2002).

4.2 Nominal stock price as a ranking criterion

Figure 3 shows the evolution of the percentage of the trace of the correlation matrix explained by the first factor in 2014 when nested PCAs are performed. The ranking criterion used to introduce new stocks (50 stocks are added at each step) in the PCAs is the nominal price of stocks. The bold (dashed) curve identifies the evolution of the first eigenvalue as a percentage of the trace when stocks are entered in the matrix in the decreasing (increasing) order of nominal price. Of course, the two curves end at the same point where all stocks are considered. The shape of the curves is qualitatively unchanged when three factors are considered. The only difference is that the upper curve starts at 40% (instead of 33%) and the lower at 15% (instead of 7%). With three factors, the two curves end at 24% (instead of 20% on figure 3). These figures show that in 2014, the first factor accounts for 20% of the trace, and factors 2 and 3 only account for 4% of the trace when all stocks are considered. We highlight that the bold curve is always above the dashed curve with one common factor, which means that the share of idiosyncratic volatility is always larger for small price stocks than for large price stocks.

Save View full size Expand inline Collapse inline

Figure 3: Percentage of trace on the first common factor as a function of the number of stocks (year 2014)

This figure presents the percentage of trace on the first axis as a function of the number of stocks for the year 2014. The nested sequence of Principal Component Analyses (PCA) starts with 50 stocks. 50 stocks are added for the next PCAs. The ranking criterion is increasing (decreasing) nominal stock price for the dashed (bold) curve.

In an efficient market, nominal prices play no role. Thus no specific shape is expected when such a ranking criterion is used, but we observe the same relationship for every single year of our 1980-2014 period. It means that if there is only one common factor, the idiosyncratic volatility of small price stocks is much higher than the idiosyncratic volatility of large price stocks.

This table shows the percentage of measure points where the idiosyncratic volatility is higher for small price stocks than for high price stocks. The two first columns give the percentages when all stocks are considered, respectively for a 1-factor and a 3-factor model. The two last columns restrict the sample to the tercile of large capitalizations.

Each year$t$, denote$N_S\left(t\right)$the number of stocks under consideration. We perform $N_t$ nested PCAs (as in Figure 3), where $N_t=N_S/50$because we start with 50 stocks in the first PCA and add 50 more stocks at each step. Denote${\lambda }_D\left(n,t\right)$ (${\lambda }_I\left(n,t\right)$) the first eigenvalue, in percentage of the trace, obtained in the n-th PCA when the decreasing (increasing) order of nominal prices is used to introduce stocks. The measure of the strength of the relationship between price and idiosyncratic volatility is denoted $\theta$ and defined in year $t$ by:

${\theta }_t=\frac{1}{N_t}{\sum }_{n=1}^N{\text{1}}_{{\lambda }_D\left(n,t\right)-{\lambda }_I\left(n,t\right)>0}$
                        (18)

${\xi }_t=\frac{1}{N_t}{\sum }_{n=1}^N\left({\lambda }_D\left(n,t\right)-{\lambda }_I\left(n,t\right)\right)$
                    (19)

where${\text{1}}_A$ is the indicator function of the event $A$ worth 1 (0) when $A$ is true(false).

${\lambda }_D\left(n,t\right)$
and${\lambda }_I\left(n,t\right)$are the percentages of variance explained by the common factor. In the three-factor version, we add the three first eigenvalues but the interpretation is unchanged.

Our results over the complete period 1980-2014 are summarized in Table 2 (page 42). The first (second) column contains the indicator${\theta }_t$when a one (three)-factor model is used. The third (fourth) column gives the same result as the first (second) column when only the tercile of large caps is used in the PCAs.

We show that, either in the 1-factor model or in the three-factor model,${\theta }_t$is always equal to 100%, which means that the idiosyncratic volatility (in percentage of total volatility) is always greater for small price stocks than for large price stocks. This result could be misinterpreted as a small-size effect because of the positive correlation between nominal stock prices and market capitalizations (Baker et al., 2009). In particular, we could argue that retail investors invest more in small firms, which may generate idiosyncratic volatility in returns of small caps. Columns 3 and 4 of Table 2 show it is not the case. For 28 years out of 35,${\theta }_t$is equal to 100% in the one and three-factor versions of the model.

Moreover, when${\theta }_t$is close to 95%, the two curves of Figure 3 cross at the point located just before the final point. The only cases deserving a special comment are 2009 in the 1-factor model and 2004 in the two models for the tercile of large caps. In 2009, the two curves (respectively decreasing and increasing order; all the graphs are available upon request) are very close. Hence, the nominal stock price is not a strong determinant of the idiosyncratic volatility in this particular year. 2009 is a year of partial recovery from the 2008 crisis with an average yearly return of 30%. This average return is, however, very different for small price stocks and large price stocks. The first column of Table 3 (page 43) shows a difference of 42% between these two average returns in favor of small price stocks. Such a difference has an impact on the ranking of nominal prices that can explain the absence of a significant difference in idiosyncratic volatilities based on the stock price. For 2004, we do not have a satisfying explanation; either in the 1-factor or the 3-factor model the two curves are almost indistinguishable when the number reaches 500. Each tercile contains approximately 1,500 firms (for a total of 4,496 firms in 2004). As a consequence, the two analyses (decreasing and increasing order of prices) with 500 firms include disjoint sets of firms. It is the reason why we do not have a satisfying interpretation for 2004.

5. Conclusion

Using the approximate factor structures defined by Chamberlain and Rotschild (1983), this paper highlights that nominal prices are a strong determinant of the idiosyncratic volatility of stock returns. In a large empirical study including approximately 8,000 stocks over 35 years, we show that small price stocks exhibit a much higher idiosyncratic volatility than high price stocks, even after controlling for the total volatility of returns. This small price effect is not a small size effect because the relationship remains valid for 34 of 35 years when we restrict the analysis to the tercile of large capitalizations. In fact, the higher idiosyncratic volatility is associated with a higher skewness, which allows us to conclude that small price stocks share some features with lotteries characterized by a small (or negative) expected return but a small probability of winning a big amount. Overall, our result is not in line with the efficient market hypothesis, which considers that nominal prices do not influence the distribution of returns.

References

• Akerlof GA. 2007. The Missing Motivation in Macroeconomics. American Economic Review 97:5-36.
• Albuquerque R. 2012. Skewness in Stock Returns: Reconciling the Evidence on Firm Versus Aggregate Returns. Review of Financial Studies 25: 1630-1673.
• Bai J, Ng S. 2002. Determining the number of factors in approx- imate factor models. Econometrica 70:191-221.
• Baker M, Bradley B, Wurgler J. 2011. Benchmarks as limits to arbitrage: Understanding the low-volatility anomaly. Financial Analysts Journal 67: 40-54.
• Baker M, Greenwood R, Wurgler J. 2009. Catering through nominal share prices. Journal of Finance 64:2559-2590.
• Birru J, Wang B. 2016. Nominal Price Illusion. Journal of Financial Economics 119:578-598.
• Campbell JY, Lettau M, Malkiel BG, Xu Y. 2001. Have Individual Stocks Become More Volatile? An Empirical Exploration of Idiosyncratic Risk. Journal of Finance 56:1-43.
• Chamberlain G, Rotschild M. 1983. Arbitrage, Factor Structure, and MeanVariance Analysis of Large Asset Markets. Econometrica 51:1281-1304.
• Connor G, Korajczyk RA.1993. A Test for the Number of Factors in an Approximate Factor Model. The Journal of Finance 48:1263-1291.
• Conroy RM, Harris RS, Benet BA. 1990. The Effects of Stock Splits on Bid-Ask Spreads. The Journal of Finance 45:1285-1295.
• Copeland TE. 1979. Liquidity Changes Following Stock Splits. The Journal of Finance 34:115-141.
• DeBondt WFM, Thaler R. 1985. Does the Stock Market Overreact?. The Journal of Finance 40:793-805.
• Doran JS, Jiang D, Peterson DR. 2011. Gambling Preference and the New Year Effect of Assets with Lottery Features. Review of Finance 16: 685–731.
• Geweke J, Zhou G. 1996. Measuring the Pricing Error of the Arbitrage Pricing Theory. Review of Financial Studies 9:557-587.
• Harvey CR, Liu Y, Zhu H. 2015. …and the Cross-Section of Expected Returns. Review of Financial Studies 29:5-68.
• Jones CS. 2001. Extracting factors from heteroskedastic asset returns. Journal of Financial Economics 62:293-325.
• Kearney C, Poti V. 2008. Have European Stocks become More Volatile? An Empirical Investigation of Idiosyncratic and Market Risk in the Euro Area. European Financial Management 14:419-444.
• Kumar A. 2009, Who gambles in the stock market. Journal of Finance 64:1889-1933.
• Todd M, Vorkink K. 2007, Equilibrium underdiversification and the preference for skewness. Review of Financial Studies 20:1255-1288.
• Patrick R, Roger T, Schatt A. 2016. Behavioral Biases in Number Processing: the Case of Analysts’ Target Prices, Working paper 6th Helsinki Finance Summit.
• Ross SA. 1976, The arbitrage theory of capital asset pricing. Journal of Economic Theory 13:341-360.
• Trzcinka C. 1986. On the Number of Factors in the Arbitrage Pricing Model. The Journal of Finance 41:347-368.
• Weld WC, Michaely R, Thaler R, Benartzi S. 2009. The nominal share price puzzle. Journal of Economic Perspectives 23 :121-142.