This post reviews the analysis in Campbell, Lettau, Malkiel, and Xu (2001) who find that firm level volatility has been rising over the period from July 1962 to December 1997. I’ve posted the code I used here. What does this mean? The authors look at the day-to-day variations in the stock returns of all publicly listed firms on the NYSE, Amex, and NASDAQ exchanges in each month, , during this sample. They then show how to decompose the variance of the daily returns each month for a typical firm (e.g., a firm selected randomly with probability proportional to its market cap) into a market-specific component, an industry-specific component, and a firm specific component. Campbell, Lettau, Malkiel, and Xu (2001) find that this firm-specific variance has been steadily rising as plotted in the figure below.
I begin by detailing how Campbell, Lettau, Malkiel, and Xu (2001) estimate their market-specific, industry-specific, and firm-specific (i.e., idiosyncratic) volatility components. A natural first approach would be to estimate industry-level regressions:
where denotes the value-weighted excess return on the market, denotes the value-weighted excess return on industry , and denotes the excess return on stock . You could then just insert the realized and terms into expressions for the variation in and to get the desired result, no?
The problem here is that and are hard to estimate and may well vary over time. Thus, a -independent procedure is necessary. After describing this procedure in Section 2, I then replicate the variance time series used in Campbell, Lettau, Malkiel, and Xu (2001) in Section 3. Finally, in Section 4 I conclude by extending the sample period to December 2012 and discussing the interpretation of the results.
The analysis in Campbell, Lettau, Malkiel, and Xu (2001) ties in closely with numerous other findings in asset pricing, macroeconomics, and behavioral finance. In the empirical asset pricing literature, Ang, Hodrick, Xing, and Zhang (2006) find a puzzling result that firms with high idiosyncratic volatility “have abysmally low average returns.” e.g., stocks with the highest idiosyncratic volatility have lower excess returns than stocks with the lowest idiosyncratic volatility. In a macroeconomic context, this analysis is directly supports the granular origins theory of Gabaix (2011) which proposes that the key source of aggregate macroeconomic fluctuations is idiosyncratic firm-specific shocks to large firms. Finally, in a behavioral finance setting, the fact that market and industry models explain so little of the variation in firm-level stock returns (somewhere between and ) suggests a new kind of problem for traders: scarce attention. “What this information consumes is rather obvious,” writes Herbert Simon. “It consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention, and a need to allocate that attention efficiently among the overabundance of information sources that might consume it.” As highlighted in Chinco (2012), it takes time to sift through all of the competing information about each and every firm and subset of firms.
2. Statistical Model
In this section I explain how Campbell, Lettau, Malkiel, and Xu (2001) compute the market-wide, industry-level, and idiosyncratic contributions to the variation in a firm’s daily returns. The challenge to doing this is that estimating the firm-specific s empirically is noisy and unreliable. To get around this challenge, Campbell, Lettau, Malkiel, and Xu (2001) decompose the daily return variance of a “typical” firm rather than every firm. e.g., they think about the determinants of the daily return variance of a firm selected at random each month from the market with probability proportional to its relative market capitalization.
To see why looking at a typical firm might be helpful here, define the relative market capitalization of an entire industry, , and the relative market capitalization of a particular firm, , in month as follows:
where and . I use the notation that denotes a particular firm in the market, and use to denote the subset of firms in industry : . The key insight is that the value weighted sums of the industry-level and firm-specific s have to sum to unity:
This observation follows mechanically from the fact that the market return is just the value-weighted sum of its industry constituents and the industry returns are just the value-weighted sums of their firm constituents:
By sampling appropriately, you can eliminate the pesky s from Equation (3) by converting their weighted sums into s.
How would you go about decomposing the variance of a typical firm in practice? First, consider estimating the market-wide and industry-specific variance components in a -independent fashion. Instead of running the regression in Equation (1), consider computing the difference between each industry’s value weighted return, , and the value weighted return on the market, :
where now lacks the tilde and is connected to via the relationship:
Since is not a regression residual, it will not be orthogonal to the market return, . Thus, when computing the the variance of the value weighted industry return, , we get:
Applying the sampling trick described above then allows us to remove the s by averaging over all industries:
This result says that if you select an industry with probability , then the expected variance of this industry’s daily returns in month will consist of a market component, , and an industry-specific component, . When the market component is big:
then most of the variation in value weighted industry returns tends to come from broad market-wide shocks. Conversely if the industry-specific component is relatively large, then most fo the variation in value weighted industry returns tends to come from different industry-specific shocks which are only felt in their particular corner of the market.
Next consider estimating the firm-specific variance component in a -independent way using the same procedure. Instead of running the regression in Equation (2), I compute the difference between each firm’s excess returns and the value weighted excess returns on its industry:
Since is not a regression residual, it will no longer be orthogonal to the value weighted industry return, , and thus:
However, the same sampling trick means that the expression for the value weighted average variance over all stocks within each industry will be -independent:
3. Trends in Volatility
I now replicate these variance measures from Campbell, Lettau, Malkiel, and Xu (2001) using daily and monthly CRSP data on NYSE, AMEX, and NASDAQ stocks over the sample period form July 1962 to December 1997. I restrict the data to include only common stocks with share prices above . The riskless rate corresponds to the day T-Bill rate. First, to compute the empirical analogue of the market variance component in Equation (15), I compute the mean and variance of daily excess returns on the value-weighted market:
where denotes the number of days in month . I plot the annualized market component of the variance of daily firm returns in the figure below.
Next, I compute the industry-specific contribution to the variance of a typical firm’s daily excess returns using the Fama and French (1997) industry classification codes. There are industries in this classification system, and I code all stocks without a specified industry as their own group leading to total industries. Clicking on the image below gives a plot of the number of firms in each of these industries.
The industry-specific contribution to the variance in daily excess returns of a typical firm is then given by:
I plot the resulting time series in the figure below.
Finally, I compute the idiosyncratic contribution to the variance of the daily excess returns of a typical firm as follows:
This is the time series I plotted in the introduction. Empirically, it seems that the overwhelming majority of the daily variation in firm-level excess returns comes from idiosyncratic shocks. One ways to quantify this statement is to look at the “model” fit each month:
where the model corresponds to a market model with industry factors. The term captures the fraction of the daily variation in firm-level excess returns that is explained by model and industry factors in each month and is consistently below . i.e., more than of the daily variation is explained by firm-specific shocks!
There are a couple of interesting take away facts from this analysis. First, the nature of the variance time series dramatically changes after December 1997 when the original sample period in Campbell, Lettau, Malkiel, and Xu (2001) ends. Specifically, the post 1997 time period is dominated by a pair of volatility spikes which are not firm-specific: the dot-com boom and the financial crisis. When compared to these events, the slow run up in the firm-specific variance component looks relatively minor.
Nevertheless, even with these large macroeconomic shocks, the majority of the variation in daily firm-level excess returns is driven by firm-specific information. e.g., even during this later period the model fit, , scarcely crosses the threshold in spite of all of the systemic risk in the market! What’s more, there is substantially cyclicality in the model fit at roughly the horizon. i.e., once every the predictive power of the value weighted market and industry factors grows and then shrinks by roughly or around -to- of its baseline. One way to interpret this finding is that there should be year cycles in the profitability of technical analysis using market-wide factors.