Question

1. Briefly explain why portfolios need to be sorted on lagged Research Spending to test return predictability.

Answer 1

Firms engage in a variety of activities. Some of these activities are straightforward, and easy to assess how they will impact firm value (e.g., maintenance capital expenditures). However, some of these activities, while crucially important for the discounted value of a firm’s future cash flows, are quite uncertain and difficult to decipher how they will ultimately impact firm value. Although hard to assess, it may still be the case that analysis of publicly available information can give substantive insight into reducing the uncertainty surrounding these actions. The activity at the heart of our investigation is investment in research and development (R&D). Given that R&D stimulates innovation and technological change, which can in turn lead to improvements in productivity, living standards, and economic output, the proper allocation of R&D investment in the economy is a critical task of the market. And yet, this task is made difficult by the fact that R&D investment is such a highly uncertain activity. Perhaps as a result of this uncertainty, R&D investment has increasingly become a market-driven activity. Although the share of R&D as a percentage of GDP has remained roughly constant (between 2-3%) since the 1960s, the composition of R&D investment in the economy has shifted dramatically, away from federal spending and toward private sector spending.1 Since the late 1980s, for example, virtually all of the increases in total R&D spending have come from the private sector. The market’s role in allocating R&D investment has become more important than ever. In this paper we demonstrate that the stock market is unable to distinguish between “good” and “bad” R&D investment, despite the fact that successful innovation is in fact predictable. We show that two firms that invest the same amount in R&D can have quite divergent, but predictably divergent, future paths. Our approach is based on the simple premise that while future outcomes associated with R&D investment are uncertain, past information about firms’ success at R&D gives us insight into their potential for future success. Our empirical strategy proceeds from the notion that past track records represent one simple way to gauge the future prospects of firms. Some firms are skilled at certain activities, and some are not, and this skill may be persistent over time. Using this idea as the starting point for our analysis, we examine the predictability of firm-level R&D investment track records for future returns and future real outcomes. We find that although R&D success is predictable, persistent, and relatively simple to compute, the market largely ignores the information embedded in past track records. Our identification of past R&D success is based on a simple framework of using a firm’s past ability in translating R&D into something the firm values. We then take this “ability” of a firm at R&D and interact it with the amount of research the firm is actually undertaking. For instance, we examine the outcomes of those firms that have been quite good at R&D and are investing heavily in R&D with firms investing identical amounts in R&D, but that have poor past track records. If the market correctly takes into account the prior track records’ implications for future success, then whether firms are optimally choosing levels of R&D or not, the market should impound relevant information regarding innovation into prices. In fact, the market could even be completely incorrect in impounding the impact of every firm’s R&D expenditures (as they do have uncertain effects on future firm value), but this would still have no implication for predictability based on past information, as the market will sometimes overvalue and sometimes undervalue this innovation. We find that the market consistently misvalues innovation in an ex-ante, predictable way. Specifically, the market does not take into account the information in firms’ past R&D abilities. Firms that have been successful in the past and that invest heavily in R&D as a percentage of sales (“GoodR&D” firms) earn substantially higher future stock returns than firms that invest identical amounts in R&D, but that have poor past track records (“BadR&D” firms). A portfolio of GoodR&D firms earns equal- and value-weighted excess returns of 135 basis points per month (t=2.76) and 122 basis points per month (t=2.61), and 4-factor alphas of 90 basis points per month (t=3.11) and 78 basis points per month (t=2.27), respectively. In contrast, the portfolio of firms with poor past track records but that invests the same amount of R&D (BadR&D) earns -15 basis points per month in 4-factor value-weighted alpha (t=0.56). The spread portfolio that takes identical high R&D-level portfolios, but exploits differences in past track records, has a 4-factor alpha of 93 basis points per month (t=2.30) or over 11% per year. Returns to the “GoodR&D” (and spread) portfolios are large and significant in the first year, and then returns remain slightly positive but basically plateau in the second and third years, with no reversal. This suggests that we are not capturing a form of overreaction, but instead that the embedded information regarding innovation that the market is misvaluing is important for fundamental firm value. Our findings add to a growing literature highlighting the market’s inability to properly value investments in R&D. On one hand, some researchers argue that investors may overestimate the benefits from R&D or simply ignore the fact that many R&D investments are not profitable (Jensen (1993)), leading to the overpricing of R&Dintensive firms. For example, Lakonishok, Shleifer, and Vishny (1994) find that growth stocks earn low future returns, while Daniel and Titman (2006) show that this growth stock underperformance is concentrated in stocks with significant “intangible” information, consistent with market overreaction to intangible information that is difficult to interpret.2 However, the recent evidence on firm-level R&D activity suggests that, if anything, the market appears to underreact to the information contained in R&D investments. For example, Chan, Lakonishok, and Sougiannis (2001) and Lev and Sougiannis (1996) demonstrate that firms with high ratios of R&D relative to market equity earn high subsequent returns; Eberhart et al. (2004) find that large increases in R&D expenditures predict positive future abnormal returns; and Hirshleifer et al. (2010) show that firm-level innovative “efficiency” (measured as patents scaled by R&D) forecasts future returns.3 We show that our results are unaffected by the inclusion of these measures in our tests, and are roughly 3 times larger in magnitude than the findings in, for example, Eberhart et al. (2004) and Hirshleifer et al. (2010), suggesting that our approach is picking up a new and previously undetected pattern in the cross-section of stock returns associated with the market’s misvaluation of high R&D ability firms. To combat the concern that our results are due to data mining, we run a series of out-of-sample tests on our findings. We find that our classification of high ability R&D firms is also predictive of future returns in an international sample (including the UK, Japan, and Germany) and in the period immediately preceding our sample period (1974-1980). For example, when we employ our baseline Fama-MacBeth cross-sectional regression on an international sample that pools together the universe of stocks from the UK, Japan, and Germany (using dollar-returns on all stocks), we find a coefficient on R&Dhigh*abilityhigh of 0.501 (t=2.24), which is similar in magnitude and significance to our U.S. findings. In addition, while our baseline U.S. portfolio results are driven by a small number of firms (the High Ability-High R&D portfolio described above contains an average of 10 stocks per month), the percentage of market capitalization in the portfolio (0.71% of the stock market’s annual value on average) is larger than that of the “small value” portfolio (0.50% of the stock market’s annual value on average) that is featured in hundreds of asset pricing papers, and which remains one of the most studied anomalies in the literature. Lastly, we run a series of tests designed to pinpoint the mechanism behind our results. First, we explore real outcomes associated with our high R&D ability firms. Specifically, we show that the firms that we classify as high ability firms and that invest heavily in R&D also produce tangible results with their research and development efforts. They generate significantly more patents, achieve significantly more patent-citations, and develop significantly more new products than firms that invest the same levels of R&D, but have poor track records. In addition, we demonstrate that high ability firms exhibit significant persistence in R&D skill, that this skill may be positively related to the presence of a founder, and that the market’s failure to understand the implications of R&D track records is related to heterogeneity in information provision by firms. For example, we show that the predictability in future returns is significantly lower for high ability firms who provide more earnings guidance; under the assumption that firms that provide more earnings guidance are also likely to provide more information to investors more generally (as in Jones (2007)), these findings suggest that cross-sectional variation in information opacity may help explain why the market fails to properly understand the information embedded in firms’ past track records.

I. Data and Summary Statistics

We combine a variety of data sources to create the sample we use in this paper. We draw monthly stock returns, shares outstanding, and volume capitalization from CRSP, and extract a host of firm-specific accounting variables, such as research and development (R&D) expenditures, sales and general administrative expenses (SG&A), book equity, etc., from Compustat. We combine these items with firm-level patent data drawn from the NBER’s U.S. Patent Citations Data File,4 segment-level product data from the Compustat Segment Data File, earnings’ guidance data from First Call and CEO founder data from Fahlenbrach’s (2009) hand-collected data and the Corporate Data Library. We draw international stock return data from Datastream and accounting data from Worldscope. We filter the datastream stock return data and identify common stocks using the procedures and suggestions outlined in Ince and Porter (2006) and Griffin, Nadari, and Kelly (2010). Table I presents summary statistics for the sample we use in this paper (Panel B), compared to the entire universe of stocks on CRSP (Panel A), over our July 1980 to December 2009 sample period. Our sample includes all NYSE, AMEX, and Nasdaq common stocks (CRSP share code 10-12) with a valid (i.e., non-missing) R&D estimate in a given year, as well as a valid estimate for the "Ability" measure that features in our analysis. The notion of “Ability” is meant to capture simply how good a firm is at turning R&D expenditures into something the firm values. We have run our tests using a number of measures of what the firm “values” and our results are robust to the various measures we have tried. The measure we show in the paper is how R&D translates into actual future sales revenue of the firm.5 One additional concern may be the horizon we use to identify the translated effect of R&D on future outcomes. As we describe below, we try to be flexible on this dimension and use up to a five-year lag in measuring the impact of past R&D expenditures on future firm outcomes. Thus, for sales (reported in the paper), we compute firm “Ability” by running rolling firm-by-firm regressions of firm-level sales growth (defined as log(Salest/Salest-1)) on lagged R&D (R&Dt-j/Salest-j; where j=1,2,3,4,5). We run separate regressions for 5 different lags of R&D (i.e., R&D from years t-1, t-2, t-3, t-4, and t-5); we then take the average of these five R&D regression coefficients as our measure of ability (regression specification shown in Table II). Again, the idea behind this measure is to isolate the extent to which a given firm successfully converts its R&D investments into future sales. We have analyzed a variety of different specifications here, and our results are robust to these permutations; for example, running a single regression for each firm of sales growth on the average of the past 5 years of R&D, and using this single coefficient as our measure of ability yields similar, and often stronger, results (we show these results in Appendix Table A5). In estimating a firm’s ability, for every firm in each year we use 8 years of past data for each firm-level regression, and we then run these regressions on a rolling basis each year using the prior 8 years of data. For each regression, we require a minimum of 6 (75%) non-missing R&D observations and that at least half the R&D observations are non-zero; otherwise, we set the slope coefficients to missing values.6 Panel B of Table I indicates that our final sample is quite similar to the overall sample of CRSP stocks. Comparing characteristic-by-characteristic, our sample does contain slightly larger stocks, with a modest growth tilt relative to the overall sample of CRSP stocks. While the stocks in our sample are slightly less levered, the price momentum, turnover, and stock volatility are nearly identical to the entire universe. Overall, the differences between the two samples appear small. Panel A of Table II presents the full-sample sample averages of the rolling firm-byfirm regression coefficients that form the basis of our ability measure. The average ability estimate is 3.33, with an average sales growth of roughly 7%, while average R&D expenditures equate to roughly 17% of sales. We then turn to some diagnostics of our Ability measure. If we are truly capturing a meaningful measure of a firm’s ability at Research and Development, we might expect to see some level of persistence in this measure (i.e., it would be odd to see firms simply jump from being classified as “good” at R&D to “poor” at R&D, and back, year after year). Panel B of Table II examines this issue by showing the annual persistence in a firm’s ability quintile assignment, for yearly lags out to 5 years. We find that firms in the highest quintile of ability remain in this same top quintile in the following year 70% of the time.7 Overall, Panel B demonstrates that there is substantial persistence in firm-level R&D ability, but that firms do transition out (on average) of the high ability category within several years.8 II. Results A. Portfolio Returns In this section we examine average returns on portfolios formed using information about both a firm’s ability and its level of R&D. We scale R&D by sales, and use threeway sorts using the same methodology of Fama and French (1996), namely: R&Dlow contains all stocks below the 30th percentile in R&D (but who have R&D greater than zero), and R&Dhigh contains all stocks above the 70th percentile in R&D. We compute firm-year ability as described earlier, using the annual average of the rolling regression coefficients of sales growth on 5 lags of R&D (scaled by sales).9 We include all NYSE, AMEX, and Nasdaq stocks from July 1978 to December 2009 with lagged share prices above $5 into these portfolios, and rebalance the portfolios yearly. We characteristically-adjust returns (as in Daniel et al. (1997)) using either 25 size/book-to-market benchmark portfolios, or 125 (5x5x5) size/book-tomarket/momentum benchmark portfolios. We also compute three- and four-factor alphas (as in Fama and French (1996), and Carhart (1997)) by running time-series regressions of excess portfolio returns on the market (MKT), size (SMB), value (HML), and momentum (UMD) factor returns. In addition to these risk adjustments, we also calculate an industry benchmark-adjusted return. If the ability measure is somehow sorting on industry (so the High Ability firms are disproportionately from one industry), we may be inadvertently sorting on an industry characteristic unrelated to our ability explanation. To combat this potential problem, each month we compute each firm’s return subtracting out its industry’s return over the same month. Thus, these industry excess returns will control for any characteristic of a firm (High or Low Ability) shared by its industry, and isolate only its abnormal returns relative to other firms in the same industry.10 Lastly, as we will compare the returns of two firms that have both been spending a large amount on R&D (but with varying abilities), we have no selection bias in terms of firms that decide to engage (or not) in R&D. This also rules out any general story that there has been an unexpected positive trend for innovative firms over the past 30 years, as that would show up in all high R&D firms. Equivalently, we compare the returns within High Ability firms, varying levels of R&D, to rule out the possibility of High Ability sorting an unobserved risk. Table III reports average stock returns for monthly portfolio sorts, and illustrates our first main result: stocks that exhibit high ability in the past and that spend a large amount on R&D (i.e., stocks in the Abilityhigh / R&Dhigh portfolio, which we will call the "GoodR&D" portfolio) outperform in the future. This result holds for both equal- and value-weight portfolio returns, and for excess returns, characteristically-adjusted returns, industry-adjusted returns, and 3- and 4-factor alphas. Further, the magnitude of this outperformance is large: Panel A shows that the GoodR&D portfolio earns 135 basis points per month (t=2.76) in equal-weight excess returns, and 122 basis points per month (t=2.61) in value-weight excess returns, which translates to 17.5% and 15.7% annually, respectively. In addition, the long-short portfolio spread (Spread) between stocks in the GoodR&D portfolio and those stocks that exhibit low ability in the past but which continue to spend a large amount on R&D (i.e., stocks in the Abilitylow / R&Dhigh portfolio, which we will call the "BadR&D" portfolio), is large and significant. For example, Panels A and B shows that the raw equal-weight spread is 73 basis points per month (t=2.61), and the raw value-weight spread is 90 basis points per month (t=2.30), which translates to 9.1% and 11.4% annually, respectively. Again this result holds for both equal- and value-weight portfolio returns, and for characteristically-adjusted returns, industry-adjusted returns, and 3- and 4-factor alphas. Note that the two components of this spread portfolio (i.e., the GoodR&D portfolio vs. the BadR&D portfolio) are very similar on other characteristics (e.g., in percentiles, the average size (0.46 vs. 0.43), bookto-market (0.31 vs. 0.38), leverage (0.26 vs. 0.25), momentum (0.56 vs. 0.53), volatility (0.53 vs. 0.49), turnover (0.72 vs. 0.69), and past R&D growth (0.65 vs. 0.69) are virtually the same for both portfolios). Panel C of Table III presents additional characteristics of these portfolios. Specifically, the four-factor loadings in Panel C suggest that the GoodR&D portfolio loads negatively on value and momentum and positively on size, meaning that the stocks in this portfolio are typically large, growth stocks with poor past returns. Meanwhile the spread portfolio has no significant loadings on any of the four factors, indicating that the returns to this portfolio do not covary with any of these well-known factors. In addition, while Panel C reveals that the High Ability-High R&D portfolio contains an average of only 10 stocks per month, the percentage of combined market capitalization in this portfolio (0.71% of the stock market’s annual value on average) is larger than that of the “small value” portfolio (0.50% of the stock market’s annual value on average) that is featured prominently in the literature. The results here are not sensitive to the particular breakpoints chosen; sorting based on quintiles or quartiles produces very similar (sometimes even a bit stronger) results. For example, the equal-weighted DGTW characteristically adjusted-spread return using 5x5 sorts is 120 basis points per month (t=2.29), while Appendix Tables A4 show that this same spread return using 4x4 sorts is 95 basis points per month (t=3.57). We have additionally tried coarser sorts, as in Appendix Table A3, where we simply split by median level of R&D. These sorts, while having less power to distinguish between R&D spending levels, again yield the same result: High Ability firms that engage in more R&D spending outperform Low Ability firms that also are above median spenders on R&D. The analogous DGTW spread portfolio returns 55 basis points per month (t=3.70). Further, the High Ability-High R&D contains an average of 29 stocks per month, with an average market capitalization of 1.93%. This is larger than the combined market capitalization of value quintile portfolios #1-3 (which together account for 1.71% [=0.50+0.49+0.72] of total market capitalization on average, and which collectively account for most (80%) of the value premium from 1963-2009).11 Lastly, we also present results in Appendix Table A9 using “conditional sorts” (as opposed to the independent sorts we use for most of the paper) which sort stocks based on Ability, and *then* by R&D within each Ability bin. This approach forces the number of stocks to be equal in each portfolio bin, and by doing so increases the number of stocks in the “High-High” portfolio. Appendix Table A9 shows that for 5x5, 4x4, and 3x3 conditional sorts, the number of stocks in this portfolio increases significantly (up to 74 stocks per month), and our results remain robustly large and significant. It is also important to note here that firms only report R&D expenses once per year, and we only calculate Ability once per year. Thus, although we report monthly returns in this table, we only rebalance our portfolios once per year. We also find virtually no reversal of the abnormal returns we document here. Panel A of Figure 1 plots the spread portfolio (GoodR&D-BadR&D) of Cumulative Abnormal Returns (CARs) following portfolio formation at time 0 through the first eighteen months, and Panel B plots the GoodR&D portfolio. Both equal- and valueweighted CARs are shown, which are the size-BM-momentum-adjusted returns each month. Returns of the spread (and GoodR&D) portfolios are large and significant in the first year (documented in Table IV), and then returns drift up slightly but basically plateau. Importantly, even continuing on into the second and third years, there is no reversal in returns. This suggests that we are not capturing a form of overreaction, but instead that the embedded information about innovation that the market is misvaluing is important for fundamental firm value. Figure 2 graphs the equal-weight yearly returns to the spread portfolio.12 Figure 2 shows that the annual returns to the strategy are fairly stable across time, and the average annual return to the spread portfolio across the 29 years in our sample is 10.8% (t=2.29). In Appendix Table A2 we also split our sample into three distinct sub-periods, and show that no single sub-period drives our results. For example, the value-weight L/S spread is 46 basis points per month (t=1.01) in the 1980s, 115 basis points per month (t=2.45) in the 1990s, and 142 basis points per month (t=1.85) in the 2000s; the equalweight equivalents are 25 basis points per month (t=0.55) in the 1980s, 121 basis points per month (t=3.67) in the 1990s, and 135 basis points per month (t=2.78) in the 2000s. Further, the annual correlation of these spread portfolio returns with the excess market return is low: 0.29 for the equal-weight, and 0.11 for the value-weight.13 In Table IV we demonstrate that simple sorts on R&D or Ability alone yield no pattern in average returns. In Panel A, we present monthly portfolio returns for quintiles based on R&D (scaled by sales).14 We group stocks with no R&D (R&Dzero) into a separate portfolio.15 Panel A indicates that excess returns across the various groups are very similar, and that the spread in returns between R&Dhigh and R&Dlow, and also between R&Dhigh and R&Dzero are small and insignificant. We also characteristicallyadjust returns (as in Daniel et al. (1997)) using 125 value-weight size/book-tomarket/momentum benchmark portfolios. Again we see no pattern in the abnormal return spreads of portfolios sorted on R&D. Note that these results are not sensitive to the particular breakpoints chosen, to the particular risk-adjustment procedure employed, or to the particular scaling variables used (except for market equity, of course, which mechanically produces a scaled price effect when used as a denominator irrespective of numerator (see Fama and French (1996)). Panel B of Table IV presents the average monthly portfolio returns associated with simple sorts on our ability measure. As with the simple sorts on R&D, these sorts on ability yield no obvious pattern in excess returns or abnormal returns, and the spread between Abilityhigh and Abilitylow is always near zero and insignificant. Lastly, the correlation of R&D and Ability is -0.04. In other words, they seem to be picking up quite different information about firms.16 In summary, the results in Tables III and IV demonstrate that our simple classification scheme, which is designed to isolate high-ability firms solely based on their past success in converting R&D into future sales, produces a large spread in future abnormal returns that is not present when looking at simple sorts on R&D, or simple sorts on ability alone. This finding highlights the fact that even though two firms may spend an equal amount on research and development, it is critical to understand the likely effectiveness of these expenditures, and that one can estimate this effectiveness by simply looking at a firm’s past experience. Thus our approach offers an ex-ante method for identifying future innovation that is likely to be successful, which we show is in fact the case in Section III.