Productivity growth and stock returns: firm- and aggregate-level analyses

ABSTRACT

A firm’s stock return is affected not only by its own productivity growth rate, but also by other firms’ productivity growth rates. We show that this spillover effect is significant and time-varying, and underlies a fallacy of composition observed in late 20th century U.S. data: stock returns and productivity growth are correlated positively in firm-level data but negatively in aggregate data. This seeming fallacy of composition reflects Schumpeterian creative destruction: a few technology winners’ stocks rise with their rising productivity while many technology losers’ stocks fall with their declining productivity. Thus, most individual firms’ stock returns correlate negatively with aggregate productivity growth. This implies that technological innovation need not be a blessing for all firms and as a result, for investors holding the market. Our findings also provide a firm-level technology innovation-based explanation of prior findings that the market return correlates negatively with aggregate earnings.

KEYWORDS:

• Technological innovation
• stock return
• heterogeneity
• productivity
• fallacy of composition

JEL CLASSIFICATION:

• G10
• O33

Acknowledgements

We thank Jay Pil Choi, Martin Dierker, Cheol S. Eun, Kewei Hou, Mark Huson, Tomohiko Inui, Bong-Soo Lee, Vikas Mehrotra, Andrei Shleifer and seminar participants at International Conference on Asia-Pacific Financial Markets, CESA Bogota, Korea University, National University of Singapore, Seoul National University Economics Department, Seoul National University Finance Department, Sogang University, Western Economic Association Conference, University of Alberta and Yonsei University Economics Department. We are also most grateful to the editor and two anonymous referees for their particularly helpful comments. Earlier version of this article is also available as NBER working paper No. 19462.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Economic profit is total revenue less total costs. Productivity growth is growth in revenues less growth in total costs. Accounting profit or earnings, differs from economic profit in subtracting accounting (rather than economic) depreciation, and in not subtracting the cost of equity capital. Economic profit associated with technological progress is alternatively characterized as an entrepreneurial rent – that is, a return to creativity.

² We define aggregate variables as weighted averages of firm-level variables throughout.

³ Our sample period ends in 2006, because the BEA and the BLS ceased reporting SIC-based industry-level deflators thereafter. The newly introduced NAICS-based industry classification was unavailable before 1987.

⁴ See section ‘Total factor productivity growth measure’ for further discussion on the construction and interpretation of TFP.

⁵ Creative destruction is induced by the frontier technology. In this article, we do not use a direct measure of the technology frontier. However, during the sample period of 1970–2006 in this article, several studies show that the aggregate TFP increased due to the advance in high TFP firms despite the increase in the dispersion of TFP among firms (Chun, Kim, and Morck 2011; Chun, Kim, and Lee 2015). This suggests that the advance in the technological frontier lead by winners is a key source of the aggregate TFP growth during the sample period.

⁶ They gauge firm’s technology proximity using patent citation-weighted R&D and associate a valuation premium with this measure. They conclude that positive spillovers raise the social return of R&D to twice its private return.

⁷ The negative correlation is not universally observed in other countries (Vivian and Jiang 2011) either, suggesting that the intensity of creative destruction might also differ across countries in a specific time period.

⁸ Using cross-country data consisting of 19 developed countries between 1900 and 2011, Ritter (2012) reports a negative cross-country correlation between real per capita GDP growth and aggregate stock returns. Based on this, he posits that the technological innovation and its ensuing effects on competition, while increasing per capita GDP and consumer welfare, need not increase aggregate shareholder wealth.

⁹ We follow Foster, Haltiwanger, and Syverson (2008), Aghion et al. (2009) and many other studies using the cost-share-based TFP index method for calculating TFP growth. This approach computes TFP growth directly, avoiding issues associated with various statistical estimation procedures. More importantly, Syverson (2011) shows this approach to be more reliable where production technologies are more flexible and heterogeneous. In contrast, the commonly used alternative approach based on production function estimation (Olley and Pakes 1996; Levinsohn and Petrin 2003), though useful in many settings, is highly problematic here because it assumes identical input trade-offs and returns to scale for all firms. The crucial importance of firm heterogeneity in this study thus necessitates the TFP index approach. Nonetheless, section ‘Robustness checks’ considers other methods of calculating TFP growth (Hall 1988; Basu and Fernald 1997) as robustness checks.

¹⁰ To correct for these problems, some researchers use citation weights (e.g. Jaffe 1986; Bloom, Schankerman, and van Reenen 2013).

¹¹ The aggregate TFP growth in Equation (2) excludes firm i to prevent spurious correlations between the TFP growth of firm i and the aggregate TFP growth.

¹² Including the lagged value of $d{\pi }_{m,t}$ in Equation (2) allows an AR(1) structure in the $d{\pi }_{m,t}$. This lets aggregate TFP growth obey an AR(1) process as well. Given this, $b_i$ captures the explanatory power of ‘unexpected’ aggregate TFP growth on firm i’s stock market return. We omit the lagged value as a robustness check and find the distributional characteristics of $b_i$ to remain qualitatively similar to that described in the figures and tables.

¹³ Our results are robust to alternative specifications. For example, to avoid any look-ahead bias, we instead use CAPM β_is estimated from the prior year’s data to calculate the abnormal return in Equation (2), and then run regressions of that form. All the results remain qualitatively the same. Replicating this procedure using other asset pricing models to calculate the abnormal return in Equation (2) yields qualitatively similar results. See section ‘Robustness checks’ for details.

¹⁴ There is one caveat with the restriction on the number of observations we use in estimating Equation (3). Jovanovic and Rousseau (2001) argue that a new technology may induce private firms to list sooner to utilize risk-tolerant equity financing. If new lists prosper on average after IPOs, we may interpret this as a positive spillover effect of technology innovation. If this is the case, we are underestimating the importance of the positive spillover effect since the restriction on the number of observations may exclude many of new lists. However, Fama and French (2004) report a lower survival rate for new lists in recent decades. This shows that not all the new entrants enjoy the positive spillover effects of technology innovations, suggesting that new lists also consist of extreme winners and extreme losers (Chun et al. 2008). This reduces the concern for underestimating the positive spillover effect by not including them. We thank an anonymous referee for pointing this out.

¹⁵ Estimating regression for each firm and counting significant coefficients fail to account for cross-firm correlations. An alternative approach, firm-level panel regressions assuming homogeneous a_i and b_i coefficients across firms and clustering by time, while imposing a different and more restrictive set of assumptions, reproduces the central findings reported in this section. See section ‘Firm-level panel regressions’ for details.

¹⁶ Rolling windows induce serial correlation in firm’s estimated coefficients in addition to the cross-firm correlations within windows (previous footnote). An alternative approach, panel regressions (section ‘Firm-level panel regressions’), is more restrictive in assuming homogeneous a_i and b_i coefficients across firms and windows, but allows two-dimensional clustering (Thompson 2011) to reflect both cross-firm and time-series non-independence. This exercise confirms the findings in this section.

¹⁷ We obtain similar figures for other estimation windows as well.

¹⁸ One sector lacks significant coefficients.

¹⁹ Summing both sides of Equation (2), weighting by w_i = firm i’s prior year-end market capitalization, yields $\sum _i{w}_i\widehat{r_i}\equiv r_{m,t}-E\left[r_{m,t}\right]=\sum _i{w}_i{a}_{i}d{\pi }_{i,t}+d{\pi }_{m,t}\sum _i{w}_i{b}_i$. This leads to Equation (4) only if $a\equiv E\left[r_{m,t}\right]+\sum _i{w}_i{a}_{i}d{\pi }_{i,t}$ is a constant within each sample period. This would follow if both $E\left[r_{m,t}\right]$ and $\sum _i{w}_i{a}_{i}d{\pi }_{i,t}$ were constant. Empirically, $E\left[r_{m,t}\right]$ need not be constant (Campbell, Lo, and MacKinlay 1997) and $\sum _i{w}_i{a}_{i}d{\pi }_{i,t}$ need not be zero – although $E\left[d{\pi }_{i,t}\right]$ is fairly close to zero (between 0.7% and 0.9% in Table 1). Nonetheless, if there is little time variation in $\sum _i{w}_i{b}_i$ within estimation windows, Equation (4) serves as a parsimonious specification. A comparison of point estimates, shown below, reveals that $b\cong {\displaystyle \sum _i^i{w}_i{b}_i}$ in corresponding estimation window, validating the assumption of a constant $a$ in each window.

²⁰ If a few very large firms had b_i > 0, a positive b might ensue despite most firms having b_i < 0. However, equally weighted and value-weighted means of the b_i exhibit similar behaviour (see especially Figure 4).

²¹ Here and throughout, we define qualitatively unchanged to mean an identical pattern of signs and significance and point estimates of roughly comparable magnitude. This specification lets aggregate TFP growth obey an AR(1) process, thereby letting b gauge the importance of plausibly ‘unexpected’ TFP growth in regressions explaining the stock market return.

²² More precisely, we estimate $a_i^{\tau }$ and $b_i^{\tau }$ for each firm i and for each estimation window $\tau$. For brevity, $\tau$ is suppressed in our notation.

²³ Kogan et al. (2012) run similar firm-level panel regressions to examine the negative spillover effect. Their aggregate innovation measure, an economic importance-weighted average of other firms’ patents, attracts a significant negative coefficient, also consistent with the negative spillover effect.

Additional information

Funding

This work was supported by the National Research Foundation of Korea [NRF-2013 S1 A3 A2053312]; Institute of Finance and Banking (Seoul National University); Institute of Management Research (Seoul National University). Morck gratefully acknowledges the support from the SSHRC and the Bank of Canada.