Technological Links and Predictable Returns

Transcription

1 Technological Links and Predictable Returns Charles M. C. Lee, Stephen Teng Sun, Rongfei Wang, and Ran Zhang ** September 10, 2017 Abstract This paper finds evidence of return predictability across technology-linked firms. Employing a classic measure of technological closeness between firms, we show that the returns of technology-linked firms have strong predictive power for focal firms returns. A long-short strategy based on this effect yields monthly alpha of 117 basis points. This effect is distinct from industry momentum, and is more pronounced for more innovative firms, firms with higher investor inattention, and firms with higher costs of arbitrage. We find a similar lead-lag relation between the earnings surprises, analyst revisions, and innovation-related activities (such as patent and citation counts) of technology-linked firms. Our results are broadly consistent with sluggish price adjustment to more nuanced technological news. JEL classification: G10, G11, G14, O30 Keywords: Technology momentum, stock returns, return predictability, patents, technological closeness, limited attention, market efficiency ** Corresponding author: Charles M.C. Lee, Joseph McDonald Professor of Accounting, Graduate School of Business, Stanford University, 655 Knight Way, Stanford, CA ; Tel: (650) ; Fax: (650) ; CLEE8@stanford.edu. Stephen Teng Sun (st@gsm.pku.edu.cn) is Assistant Professor of Economics at the Guanghua School of Management, Peking University. Rongfei Wang (wangrongfei@pku.edu.cn) is a PhD Candidate in Accounting at the Guanghua School of Management, Peking University. Ran Zhang (RZHANG@gsm.pku.edu.cn) is Associate Professor of Accounting at the Guanghua School of Management, Peking University.

2 Technological Links and Predictable Returns September 10, 2017 Abstract This paper finds evidence of return predictability across technology-linked firms. Employing a classic measure of technological closeness between firms, we show that the returns of technology-linked firms have strong predictive power for focal firms returns. A long-short strategy based on this effect yields monthly alpha of 117 basis points. This effect is distinct from industry momentum, and is more pronounced for more innovative firms, firms with higher investor inattention, and firms with higher costs of arbitrage. We find a similar lead-lag relation between the earnings surprises, analyst revisions, and innovation-related activities (such as patent and citation counts) of technology-linked firms. Our results are broadly consistent with sluggish price adjustment to more nuanced technological news. JEL classification: G10, G11, G14, O30 Keywords: Technology momentum, stock returns, return predictability, patents, technological closeness, limited attention, market efficiency 1

3 1. Introduction The valuation of firms technological capabilities is becoming increasingly important for investors as society becomes more innovation-driven. For many firms today, technological prowess is an important determinant not only of short-term profitability but also of long-term survival. At the same time, a firm s technological capabilities are notoriously difficult to measure. Moreover, a growing literature reports that investors do not seem particularly good at valuing these capabilities. For example, investors tend to misvalue innovation (Cohen, Diether and Malloy, 2013) or undervalue innovative efficiency and originality (Hirshleifer, Hsu, and Li, 2013, 2017). In this paper, we find that investors also seem to overlook another potentially important source of technology-related information: recent price-relevant news affecting other technologically linked firms. Firms do not conduct its technological research in isolation; frequently, they interact with each other intensively, leading to an innovation process characterized by knowledge spillovers (Jaffe, Trajtenberg, and Henderson, 1993). Given the particular nature of innovation and its ever-growing role in firm behavior and valuation, it is important that we go beyond firms own innovation characteristics and examine the implications of firms interactions in innovation. Firms working on areas of innovation that substantially overlap with each other could be subject to similar input or output linkages, which become important transmission channels for common price shocks (Acemoglu et al., 2012). Firms with similar technologies can also benefit from the spillover effect of each other s innovation activity along 2

4 technological lines (Bloom, Schankerman, and Van Reenen, 2013). Specifically, firms working on similar technologies may use similar inputs of production, with the inputs here being broadly interpreted as anything required in the production process, i.e., valuable human resources, key raw materials, production and management equipment and structures including Information and Communication Technology (ICT) or intangible knowledge in production. For example, breakthroughs in production technology led to dramatic cost reductions in silicon chips, which in turn greatly impacted on the vitality of the electronics industry that relies on these chips as a raw material. Similarly, technological progress in touch screen technology today bodes well for the firms making products that use these touch-screens. In this paper, we examine how shocks to one firm affect other firms that are closely related in terms of their technological expertise. While previous studies document rich asset pricing implications of several types of relationship between firms, including product market link, customer-supplier link, geographical link, labor market link, and alliance link (Hou, 2007; Cohen and Frazzini, 2008; Cohen and Lou, 2012; Li, Richardson, and Tuna, 2014; Huang, 2015; Lee, Ma, and Wang, 2015; Li, 2015; Cao, Chordia, and Lin, 2016), relatively little work has been done on the pricing implications of technological affinity. If investors understand and take into account the ex ante publicly available technological links, prices of the focal firm should fully adjust when news about its technology-linked firms arrives at the market. On the other hand, if investors are slow to understand and/or do not pay enough attention to the news affecting firms that are closely-aligned in technological space, 3

5 stock prices of focal firms will exhibit a predictable lag with respect to recent news affecting its technology-linked peers. Therefore, one asset pricing implication of investors limited processing capability and attention with respect to technological links is that price movements across linked firms are predictable: specifically, focal firm prices will adjust with a lag to shocks experienced by linked firms. To better illustrate our idea that news about one firm could translate into other firms in the technology space, consider the celebrated Steve Jobs Patent (patent number 7,479,949, with Steve Jobs as Co-inventors), which was granted on January 20, This patent, titled Touch screen device, method, and graphical user interface for determining commands by applying heuristics, is the core patent of Apple s multi-touch technology, which is widely used today in iphones and ipads. The granting of this patent was accompanied by extensive media coverage right after the grant date and in subsequent years. 1 During the [t, t+2] window of the patent grant date, the abnormal return of Apple is 10.56%, which largely accounted for the firm s abnormal returns for the entire month of January. 2 Although Steve Jobs and Apple are the most direct beneficiaries of this patent grant, the event is also significant for the broader field of multi-touch technology. This is because multi-touch technology has wide applicability to many products beyond iphones and ipads. 3 For example, touch screens can be used in products 1 For example: The wide influence of the Steve Jobs Patent is also reflected in the enormous number of forward citations: as of August 31, 2017, it has been cited more than 1,000 times, according to the google patent website. 2 The 3-day abnormal return is defined as 3-day cumulative market adjusted return, where the market return is the CRSP value-weighted average return. 3 As the title suggests, multi-touch technology allows two or more fingers to be used on the touchscreen, and this enables devices to recognize and respond to more than one touch at the same time. 4

6 such as televisions, interactive screens, and ATMs. This technology also has application in contexts as wide as in-car instrument panels, intelligent home appliances, hospitality counters, among others. According to Research and Markets, a market research company, the global market for multi-touch screen alone was valued at $6 billion in 2016, and is projected to reach $16 billion by Therefore, while the granting of this patent confers some monopoly power upon Apple, it also is an endorsement of the potential value and the technological feasibility of multi-touch technology as a whole. It seems likely that other firms with similar technological expertise will also benefit from this new information. In Appendix Table A1, we report the returns for a portfolio of firms that are most closely linked to Apple in terms of their technology, in the months surrounding the approval of the Steve Jobs patent. Market-adjusted return (MAR) is defined as raw monthly return minus CRSP value-weighted monthly return, and all stocks within a given portfolio are weighted by a measure of technology closeness following Jaffe (1986). For the month of the patent grant (January 2009), the full-sample MAR is 6.43% (t=6.10). In the subsequent month (February), the full-sample MAR is 2.67% (t=4.71). These results suggest some contemporaneous stock comovement, as well as some lead-lag effect. Interestingly, the effect of the contemporaneous comovement is stronger for firms in the same 2-digit industry as Apple, while the lead-lag return predictability is more pronounced for firms in different industries. In fact, we find that technological proximity is directly related to the intensity of these The multi-touch technology greatly expands the range of functionality that devices can support. For instance, pinch to zoom is a classic function that works with multi-touch technology

7 results both contemporaneous comovement and lead-lag return predictability are stronger in the portfolio with the closest technology-link. These anecdotal findings suggest that the pricing implications are stronger for firms with closer technological ties. 5 The Steve Jobs example demonstrates that technological affinity, an idea that was first documented in Jaffe (1986), can capture an important dimension of inter-firm relationship. To the extent that investors do not immediately recognize the price relevance of this information, we would expect a diffused price adjustment process along the lines of these technological links. To test for the return predictability of technology-linked firms more generally, we implement the following portfolio strategy. For each focal firm i, we calculate the weighted return of a portfolio of firms that share similar technology as the focal firm, TECHRET i = j i TECH ij RET j / j i TECH ij where RET j is the return of firm j and TECH ij measures the degree of technology closeness between firm i and j. Specifically, we adopt the classic approach first pioneered in Jaffe (1986) and developed in Bloom, Schankerman, and Van Reenen (2013) to calculate TECH ij, which exploits firm-level data on patent distribution out of 427 different technology classes to locate firms in a multidimensional technology space. We then sort focal firms into decile portfolios based on lagged returns of their technology-linked firms, and find strong evidence that lagged returns of those technology-linked firms have significant return predictability for focal firms. Specifically, a portfolio that goes long in those focal firms whose technology-linked 5 This results also indicate that it is important to use the degree of technology closeness as weights to calculate technology-linked returns (TECHRET). 6

8 firms performed best in the prior month and goes short in those focal firms whose technology-linked firms performed worst in the prior month, yields an equal-weighted return of 117 basis points (t=5.47) per month. For the analogous value-weighted portfolio, the returns are 69 basis points per month (t=3.19). We refer to this return predictability as technology momentum. In subsequent tests, we show these return prediction results are robust to a variety of controls, including size, book-to-market, gross profitability, asset growth, R&D intensity, short-term reversal and medium-term price momentum variables. It is natural for firms in the same industry to share similar technologies, so perhaps the return predictability we document is a rediscovery of the well-known industry momentum effect (Moskowitz and Grinblatt, 1999; Hou, 2007). On the other hand, prior studies report that while firms technology space has some overlap with their product market space, firms in distinct product markets or industries also often invest in similar technologies. 6 Empirically, we also find that firms in the closest technology-linked cohorts can come from many different industries. 7 To formally test the prediction that the technology momentum effect that we 6 For example, Bloom, Schankerman, and Van Reenen (2013) document that while IBM, Apple, Motorola, and Intel are all close in technology space, as demonstrated in their patenting activity and joint research partnerships, but only IBM and Apple compete in the same PC market and only Motorola and Intel compete in the semiconductor market. They also document firms competing in the same industry (or more specifically, the same product market) may invest R&D in distinct technologies. For instance, Gillette and Valance both compete in batteries but Gillette does R&D mainly in personal care products while Valance developed a new phosphate technology for lithium ion batteries. 7 In our tests, an average (median) patent technology class contains firms from 10 (10) different 2-digit SIC industries and 31 (26) different 4-digit SIC industries. The average HHI concentration ratio in a patent technology class by different 2-digit (4-digit) industry is 0.37 (0.21), showing that various industries indeed secure patents in the same technology class. An HHI concentration ratio close to 1 in our case means the majority of patents in one technology class comes from one industry while a smaller ratio means patents are more evenly distributed among different industries in a given technology class. In comparison, the HHI concentration ratio based on sales for an average 3-digit SIC industry in Hou and Robinson (2006) is

9 documented is not driven by the industry momentum effect identified by Moskowitz and Grinblatt (1999), we incorporate past and current industry returns in our tests and find that the technology-linked results are even stronger in the presence of industry controls. In further tests, we also control for supplier and customer returns, and pseudo-conglomerate returns, and find that our main results continue to hold. Overall, these tests show that the technology momentum that we documented is distinct from return momentum arising from industry links, customer-supplier links, and standalone-conglomerate firm links. To establish the robustness of this return predictability result, we conduct a series of additional tests. First, we examine the predictive relation in sample sub-periods. Dividing the full sample into four sub-periods, we find a clear and robust lead-lag return relation in every sub-period. Next, we examine the sensitivity of our results to the age of the technological closeness measure. Specifically we compute the closeness measure based on patent issuance data that is available in year t, t-1, t-2, and t-3. Our results show that measures of TECHRET relying on lagged one-, two-, and even three-year technology closeness data still significantly predict focal firm returns, although the predictive power decreases as the technology mapping becomes more stale. Evidently, investors can form trading strategies like ours even with relatively old measures of technological closeness. Finally, we perturb the length of the return estimation and holding periods. Following Moskowitz and Greenblatt (1999), we use various (L, H) strategies, whereby the technology momentum portfolios are formed on L-month lagged returns, 8

10 held for H months, and rebalanced monthly. We examine various different lags (L = 1, 3, 6, and 12) and a variety of different holding periods (H = 1, 6, 12, 24, 36). We find that the profitability of shorter-term strategies is not sensitive to the length of ranking period L. The equal-weighted raw monthly return for (1,1) strategy and (12,1) strategy is 1.17% and 1.11%, respectively. We also find that the profits decay monotonically and diminish to be insignificant in the longer holding period H. Specifically, we observe no sign of any return reversal in the longer period, indicating that the return predictability of technology momentum cannot be explained by investors overreaction. Having established the robustness of the main result, we then conduct a series of cross-sectional tests to shed light on how this result varies across different firms and types of news. Ex ante, we posit that the magnitude of the delayed price reaction will be an increasing function of: (a) the relevance of technology-linked firm news to the focal firm, (b) the extent to which investors are inattentive, and (c) the relative costs of arbitrage. Specifically, we expect a stronger effect for focal firms that are more innovation-driven (as measured by the size of their R&D spending and patent-related activities, both scaled by book equity). If the technology momentum we document represents a market inefficiency driven by investors limited attention and information processing capacity, we should find a stronger effect for firms that investors are more likely to overlook. To test this hypothesis, we use firm size, analyst coverage, and institutional ownership as proxies for investor attention. Finally, if the return 9

11 predictability reflects mispricing, we would expect to see a greater effect in situations where arbitrage costs are higher (firms with greater idiosyncratic volatility, or in the case of bad news). 8 Our test results support all these conjectures. We find strong evidence that the technology-linked momentum effect is more pronounced when the focal firm is: smaller, has fewer analysts covering them, has lower institutional ownership, and exhibits higher arbitrage costs (proxied by idiosyncratic volatility and bad news). The effect is also much stronger (more than doubled) for firms with higher than median R&D spending and patent-related activities. These results further support the view that technology momentum is a mispricing phenomenon driven by investors limited attention, valuation uncertainty, and arbitrage costs. Finally, we conduct a series of tests designed to establish the fundamental nature of the lead-lag pattern observed in returns. An alternative to the mispricing explanation is that firms with similar technology are exposed to similar risk factors, and that variations in these risk factors are driving the lead-lag return pattern. It is not easy to think of examples of risk factors that might behave in ways that give rise to these patterns. Nevertheless, a risk-based argument is difficult to rule out using evidence based on stock return correlations alone. In our final set of tests, we turn to measures of real activity and show that technological linkages are helpful in identifying lead-lag correlation in the operating performance as well as the innovation-related activities of tech-linked firms. 8 Prior studies suggest bad news tend to be incorporated into price more slowly (Hong, Lim, and Stein, 2000), either because investors are more reluctant to sell their losers, or because short-selling is more costly to implement (Beneish, Lee, and Nichols, 2015). 10

12 Specifically, we find that the lagged earnings surprises (SUEs) of the technology-linked firms are strong predictors of focal firm earnings surprises (SUEs), even after controlling SUEs of the focal firm from the last four quarters. In fact, the SUEs of the tech-linked firms have predictive power for the focal firm s SUE over the next four quarters. We observe a similar pattern in analyst forecast revisions (FREV). Specifically, we find that, after including a host of control variables, lagged FREVs of the tech-linked firms have strong predictive power for the FREV of the focal firm. These earnings-based results strongly suggest that the return patterns we documented earlier have their root in lead-lag patterns in firm fundamentals. Furthermore, we find a similar lead-lag pattern in two important innovation-related activities: the patent and citation counts. Specifically, we show that the average number of patents granted in a given year to a portfolio of technology-linked firms (where each peer firm is weighted by its pairwise closeness to the focal firm) has significant predictive power for the patent applications of the focal firm in the subsequent year. Similarly, the citation counts of technology-linked peers (the number of forward life-time citations received by patents granted in a given year) is a significant predictor of the citation counts of the focal firm. These results are robust to the inclusion of a myriad of control variables, including year and industry (or firm) fixed effects. Taken together, the patent and citation counts results document a strong innovation spillover effect along technology-linked firms, which lends further credence to the sluggish price adjustment hypothesis. The remainder of the paper is organized as follows. Section 2 lays out the 11

13 background for the setting we examine in the paper. Section 3 describes the data and variables. Section 4 provides our main results on technology momentum. Section 5 conducts more extensive robustness tests while Section 6 examines the mechanisms in more detail. Section 7 explores the real effects of the technological link and Section 8 concludes. 2. Background Our paper is broadly related to several strands of literature. Firstly, there is a rich literature documenting the patterns and consequences of investors limited attention to information with substantial value implications. Theoretical works starting with Merton (1987) examine the effect of investor inattention in security prices, followed by later studies including Hong and Stein (1999), Hirshleifer and Teoh (2003), and Peng and Xiong (2006). The general message from these models is that delayed information recognition due to investors limited attention can give rise to return predictability, beyond explanations by traditional asset pricing models. A growing empirical literature is lending substantial support for these models predictions. 9 In particular, Cohen and Frazzini (2008) find investors pay limited attention to the performance of focal firm s economically linked firms, i.e., the customer firms. Consequently, focal firm s stock price does not immediately incorporate news involving linked firms, generating predictable future price movement. We here study a more nuanced but nevertheless important link between 9 Exemplary works include Huberman and Regev (2001), Barber and Odean (2007), DellaVigna and Pollet (2009), Hou (2007), Menzly and Ozbas (2010), and Hong, Torous and Valkanov (2007). 12

14 firms: their distance in technology space. Given limited investor attention, we posit the value implications of this link will only be fully priced gradually over time, particularly for firms that are more costly to arbitrage. 10 Secondly, our work relates to research that examines investors limited information processing capacity and its ramifications. Investors biased interpretation of information could lead to a significant delay in the impounding of information into asset prices. Tversky and Kahneman (1974) and Daniel, Hirshleifer, and Subrahmanyam (1998), among others, show that this bias could stem from investors overweighting their own prior beliefs and underweighting observable public signals. Numerous recent empirical works lend support to this view. For example, investors under-react to public announcements of corporate events (Kadiyala and Rau, 2004), stock splits (Ikenberry and Ramnath, 2002), and goodwill write-offs (Hirschey and Richardson, 2003). Some of this work shows industry peer firms stock return portends the focal firm s stock return. For instance, Hou (2007) finds a lead-lag pattern between weekly returns of large firms and small firms from the same industry, and Jiang, Qian and Yao (2016) find industry leaders R&D growth could have predictability for stock returns of other firms in the same industry, due to R&D spillover effects. Our study of technology-linked momentum is a context where investors are subject to limited information processing capacity along the technological links, which often transcend industry links. Thirdly, our work joins a burgeoning literature that studies the asset pricing 10 We use idiosyncratic volatility as a proxy for cross-sectional differences in arbitrage costs. Firms with greater idiosyncratic volatility are generally more costly for investors to trade (see Baker and Wurgler, 2006). 13

15 implications of firms innovation-related activities. Existing works find various aspects of firm s innovation activity, such as R&D intensity (Chan, Lakonishok, and Sougiannis, 2001), R&D growth (Penman and Zhang, 2002; Eberhart, Maxwell, and Siddique, 2004; Lev, Sarath, and Sougiannis, 2005), patent citations (Gu, 2005; Matolcsy and Wyatt, 2008), and more recently, innovative efficiency (Hirshleifer, Li and Hsu, 2013) and innovative originality (Hirshleifer, Li and Hsu, 2017) have strong predictability for its future operating performance and stock return. Our work is distinct from the existing literature in that we study the effect innovations by technologically related peer firms, rather than innovations at the focal firm itself. Our work is inspired by Bloom, Schankerman, and Van Reenen (2013), who demonstrate that a firm s technology space and product market space provide two distinct inter-firm networks, as two firms in different product market industries could produce patent in the same technology class and share similar innovative knowledge. 11 Utilizing the measure of technology closeness originally from Jaffe (1986), they demonstrate the effect of peer firms R&D spending could be decomposed into technology spillover effect (on technology space) and product market stealing effect (on product market space). Our paper also makes use of this approach to parametrize the distance between two firms in technology space and complements the literature that has focused exclusively on the product market space. In this dimension, our paper belongs to a growing accounting and finance literature 11 It is useful to note that the Bloom, Schankerman, and Van Reenen (2013) capture the effects of concurrent R&D spending, our return prediction captures the information contained in other tech peers both existing and new knowledge stock. For example, stock return of a technology peer firm could reflect investor valuation of new patent, or about valuation change to certain technology embedded in existing patents due to news reflecting potential new usage or news about new innovations that will make the existing technology obsolete. 14

16 that studies the wide implications of previously understudied firms technological link, for example, corporate cash-holding (Qiu and Wan, 2014), M&A (Bena and Li, 2014), bankruptcy (Qiu, Wang and Wang, 2016), strategic alliances (Li, Qiu and Wang, 2017) and analysts coverage and forecast (Tan, Wang and Yao, 2016). Particularly relevant to our work, Tan, Wang and Yao (2016) provide strong evidence that an information spillover effect exists between two technologically related firms and it has substantial implications for analyst coverage. Lastly, our work is related to a concurrent working paper by Bekkerman and Khimich (2017; hereafter BK) that also examines the pricing implications of firms technological link. We became aware of BK s work only as we were wrapping up our own. The motivating research question and main results of the two studies are similar. However, our paper differs from BK in several important respects. First, BK apply textual analysis to patent documents to determine technological affinity, while our paper measures pairwise distance using patent class distribution (Jaffe, 1986). 12 One advantage of our measure is that it measures the degree of technology closeness between firms, providing an economically meaningful weighting scheme to construct a weighted-average return of technology-linked firms. Second, due to more stringent data requirements, BK only examine stock returns from 1997 onward, while our approach enables us to study a much longer time period (from 1963 onward). Finally, we provide an extensive set of test results that document a lead-lag 12 The Jaffe (1986) approach is gaining wide acceptance in economics. A growing empirical literature in accounting and finance has also utilized the this approach to measure the distance between firms in technology space (such as Bena and Li, 2014; Qiu and Wan, 2015; Qiu, Wang and Wang, 2016; Li, Qiu, and Wang, 2016; Tan, Wang, and Yao, 2016). None of these studies focus on the issue of lead-lag patterns in returns. 15

17 relationship in the fundamentals of technology-linked firms (i.e., unexpected earnings, analyst forecast revisions, patent and citation counts). Overall, the two papers corroborate well with each other, and provide complementary evidence that firms technological links contain valuable information that market prices only fully incorporate gradually over time. 3. Data and Variables The main data set used in this study is the Google patent data generously provided by Kogan et al. (2017). 13 Specifically, Kogan et al. (2017) use Optical Character Recognition (OCR) technology and a number of textual analysis algorithms to extract relevant information from the patent document, and then map the identified assignees to the Center for Research in Security Prices (CRSP) unique identifiers (PERMNO). This dataset covers 1.9 million CRSP matched patents granted by the US Patent and Trademark Office (USPTO) from 1926 to We extract CRSP matched patent information from the Google patent data to construct our technology-linked variables. Since we focus on the stock market implication for the patent information, it is important to identify patents that are publicly available to investors to avoid look-ahead bias. In particular, there are two important time points for each patent: the application date and the grant date. The application date is the date that the 13 The Google patent data are available at: 14 The Google patent data has a more extensive coverage than the NBER patent data developed by Hall, Jaffe, and Trajtenberg (2001). For example, during the same period covered by the NBER patent data ( ), the Google patent data adds an average of 2,187 patents per year to the NBER patent data and corrects some errors. 16

18 inventors apply for a new patent to the USPTO, the grant date is the date that the patent gets formally issued by the USPTO, and the lag between the two dates is on average two to three years. Unless there is a federal holiday, the USPTO issues patents every Tuesday, and its publication, Official Gazette, lists detail information of the patents granted on that day. 15 Thus the investors are able to get the patent information freely from the patent offices on the grant date. 16 Examining abnormal stock turnover around patent grant date, Kogan et al. (2017) provide evidence that patent grant conveys important information to the market and is reflected in the stock price. Therefore, we use the grant date as our key time point to identify the patent information. Our main sample consists of firms in the intersection of the Google patent data, CRSP and COMPUSTAT. We focus the analysis on common stocks (CRSP share codes 10 and 11) and exclude financial firms with one-digit SIC codes of six. To insure that the relevant accounting and patent information is publicly known to investors in the market, we impose at least a six-month gap between fiscal-year end month and stock returns in our stock return tests. Specifically, we first match the Google patent data for grant year t with COMPUSTAT accounting data for the same fiscal year t, and then match to CRSP stock returns data from July year t+1 to June year t+2, as in Fama and French (1992). We require firms to have non-missing 15 The USPTO patent information is available at: 16 We note that after the American Inventors Protection Act (AIPA), which came into effect on November 30, 2000, the USPTO began publishing patent applications 18 months after the application date. For patents that were filed after November 30, 2000, the market had full knowledge of the patent at the publication date (which is 18 months after the application date) or the grant date, depending on which is earlier. In contrast, for the patents filed before November 30, 2000, the grant date was the earliest time for the market to know the patent. 17

19 market equity and SIC classification code from CRSP, and non-negative book equity data at the end of previous fiscal year from COMPUSTAT. We further restrict our sample to firms that have at least one patent granted in the rolling-window of past five years. 17 In order to reduce the impact of micro-cap stocks, we exclude from our sample stocks that are priced below one dollar a share at the beginning of the holding period. 18 Moreover, we employ the return correction approach suggested in Shumway (1997) for the delisting bias, though these adjustments have no effect on our results. In addition to stock returns, we obtain institutional holdings data from the Thomson Reuters 13F dataset, and analyst forecast data from Institutional Brokers Estimate System (IBES) unadjusted files. Specifically, in each month we get most recent mean consensus forecasts as well as the analyst coverage number immediately prior to the portfolio formation date. Following Jegadeesh et al. (2004), analyst forecast revision is defined as the change of one-year-ahead earnings consensus forecasts for the same fiscal year in a given month, scaled by the beginning price of the month. We measure technology-linked return (TECHRET) as the average monthly return of technology-linked firms in the technology space weighted by pairwise technology closeness, which is determined on the basis of 427 different technology classes defined by USPTO for all patents. Formally, technology-linked return for firm i and 17 In the robustness tests (Panel A, Online Appendix Table A.1), we further require the sample firms to have at least two or three years with granted patents in the rolling-window of past five years, and our results are robust with this alteration. 18 In the robustness tests (Panel B, Online Appendix Table A.1), we exclude stocks with price below five dollars a share or market capitalizations below the 10th percentile of NYSE stocks in our analysis, and find results unchanged. 18

20 month t is defined as: TECHRET it = j i TECH ij RET jk / TECH ij j i where k ={1, 3, 6, 12} represents the number of past month returns of technology-linked firm j. Unless otherwise noted, we use k = 1 to construct our main variable throughout the paper. We also take k = 3, 6, 12 in our robustness tests. Following Jaffe (1986) and Bloom, Schankerman, and Van Reenen (2013), TECH ij is the technology closeness defined as the uncentered correlation between all pairs, (T i T j ) TECH ij = (T i T i ) 1/2 (T j T j ) 1/2 where T i = (s 1, s 2,, s τ,, s 427 ) is the vector of firm i s technology activity of 427 elements and the τ th element s τ is the average share of number of patents in USPTO technology class τ out of the firm i s total number of patents over the rolling past five years. Technology closeness ranges between zero and one, depending on the degree of overlap in technology space, and is symmetric in firm ordering (i.e., TECH ij = TECH ji ). Note that by construction the TECH essentially acts as a weight in calculating the average stock return of technology-linked firms and is biased toward firms more technologically close to the focal firm (i.e., firms more adjacent to the focal firm in technology space is given higher weights). TECH is calculated at the end of each year t based on patent grant date that is publicly available, and then mapped to the return data from July year t+1 to June year t+2. Other variables are defined in Appendix Table A2. The final sample consists of 561,989 firm-month observations spanning July 1963 to June 2012 (i.e., 588 months). Panel A of Table 1 presents the descriptive 19

21 statistics of sample firms. The number of firms varies from a low of 189 firms in July 1963 and a high of 1,363 firms in June The sample firms cover almost 53% of the CRSP common stock universe in terms of market capitalization. This is unsurprising since we only include the sample firms with at least one patent granted in the past five years. We note that the average number of linked firms in the technology space is 280, and the pairwise technology closeness (TECH) has the average of 0.11 and standard deviation of 0.16, indicating that the technological link is quite pervasive and sparse. 19 Panel A of Table 1 reports summary statistics for our key variables. The average technology-linked return (TECHRET) is The distribution pattern is quite similar to industry momentum return (INDRET), and less volatile than firm past one-month return (REV). In Panel B of Table 1, several correlation coefficients are noteworthy. The Pearson correlation between TECHRETt-1 and RETt is 0.028, providing raw evidence for the lead-lag effect along the technological link. Although TECHRETt-1 exhibits trivial correlations with a bunch of traditional return predictors (i.e., size, book-to-market, gross profitability, asset growth, R&D intensity), it is considerably positively correlated with industry momentum return (INDRETt-1), past one-month return (REV) and medium-term momentum (MOM) (Pearson correlations are for INDRETt-1, for REV, and for MOM). In the subsequent analysis, we show the return predictability of TECHRETt-1 remains after controlling for other variables under various settings. 19 In the robustness tests (Panel C, Online Appendix Table A.1), we only include sample firms of which the technology-linked firms have TECH larger than 0.01 or rank in the top 50 in terms of TECH, our main results are unchanged. 20

22 4. Empirical Results 4.1. Portfolio tests Table 2 shows the basic results of our paper. At the beginning of each month, we sort all firms into deciles based on the return of their technology-linked portfolios in the previous month. The decile portfolios are then rebalanced at the beginning of each month to maintain either equal or value weights. The bottom lines show the returns of a zero-cost portfolio that hold the top 10% high technology-linked firm return stocks and sells short the bottom 10% low technology-linked firm return stocks. In Table 2, we find strong evidence consistent with technology-linked firm returns predict focal firm returns. Specifically, taking the strategy of going long in firms whose technology-linked firms performed best in the prior months and selling short those firms whose technology-linked firms performed worst (L/S), yields equal-weighted returns of 117 basis points per month (t = 5.47), or roughly 14.1% per year. The corresponding value-weighted returns from the L/S portfolio are 69 basis points per month (t = 3.19), or about 8.3% per year. In the next five columns, we control for other known return determinants. The technology-momentum strategy delivers CAPM abnormal returns of 1.22% (0.74%) per month for the equal (value) weighted portfolios. The strategy delivers Fama and French (1993) abnormal returns of 1.26% (0.80%) per month for the equal (value) weighted portfolios. Adjusting returns for the stock s own price momentum by augmenting the factor model with 21

23 Carhart s (1997) momentum factor has the only negative but negligible effect on the results. Subsequent to the portfolio formation, the baseline long-short portfolio earns an abnormal return of 1.08% (0.65%) per months for equal (value) weighted portfolios. Lastly, we adjust returns using the Fama and French (2015) five-factor model and using the five-factor model plus the momentum factor. Those adjustments have little effect on the results: subsequent to the portfolio formation, the baseline zero-cost portfolio earns abnormal returns of 1.37% (0.86%) and 1.21% (0.73%) for equal (value) weighted portfolios. The results show that after controlling for common risk factors, high (low) technology-momentum stocks earn high (low) subsequent (risk-adjusted) returns Regression results In this section, we formally test our hypothesis in a regression framework, controlling for other determinants of firm return and isolate the marginal effect of our main variable, lagged technology-linked returns. Specifically, in Table 3, we conduct forecasting regressions of focal firm returns using Fama and MacBeth (1973) regressions. The dependent variable in columns 1-3 is the focal firm return in month t (RETt). The independent variable of interest is the return of the focal firm s technology-linked firms in month t-1 (TECHRETt-1). We also include the 20 In Panel D of Online Appendix Table A.1, we show the portfolio results using alternative weighting schemes. Specifically, for the equal-weighted scheme (i.e., constructing TECHRET by giving all technology-linked firms equal weights), the long-short portfolio earns raw returns of 1.07% (0.64%) per month for equal (value) weighted portfolios. For the value-weighted scheme (i.e., weighting returns of technology-linked firms by their market capitalization), the long-short portfolio return decrease to 0.39% (0.36%) per month for equal (value) weighted portfolios. These portfolio returns are smaller than the TECH-weighted (i.e., weighting returns of technology-linked firms by the degree of technology closeness to focal firms) baseline results in Table 2, supporting the value of technology closeness (TECH) weighting scheme. 22

24 value-weighted industry return of the focal firm in month t-1 (INDRETt-1) as an independent variable, following Cohen and Lou (2012) and Moskowitz and Grinblatt (1999). Other control variables include lagged size, book-to-market, gross profitability, asset growth, R&D intensity. Lastly, we include REV, a short-term return reversal variable, defined as the focal firm s stock return in month t-1, to control for the short term reversal effect of Jegadeesh and Titman (1993), and MOM, a medium-term price momentum variable, defined as the focal firm s stock return for the last 12 months except for the past one month, to control for the momentum effect of Chan, Jegadeesh, and Lakonishok (1996). Cross-sectional regressions are run every calendar month and the time-series standard errors are Newey-West adjusted (up to 12 lags) for heteroskedasticity and autocorrelation. Table 3 column 1-3 report the basic results. Consistent with the portfolio results, TECHRETt-1 is a strong predictor of next month s technology-linked focal firm return in all three specifications. Specifically, before controlling for any other variables, the coefficient of TECHRETt-1 in column 1 is with a t-statistic of 4.18, indicating that the average monthly return spread of the focal firms in the top and bottom technology-linked return deciles is 64.3 basis point. In column 2, we include size, book-to-market, gross profitability, asset growth, R&D intensity, reversal, and momentum as control variables. The magnitude and significance of the coefficient of TECHRETt-1 are almost the same as in column 1. In column 3, we further include lagged industry return as a control variable. Both the magnitude and the significance of the coefficient of TECHRETt-1 are getting even more pronounced to predict focal 23

25 firm s return for month t after adding the industry momentum control variable, which indicates that the technology momentum effect we documented cannot be explained by the industry momentum effect documented by Hou (2007). To better distinguish our technology-momentum effect from the previously known industry momentum effect, we further report results of taking the industry-adjusted return (calculated as the difference between a focal firm s return this month and its contemporaneous industry return) instead of the raw return as a dependent variable. By subtracting the industry return from focal firm return, we purge out stock return continuation that arises from industry wide return auto-correlations. Column 4 of Table 3 indicates that, even after subtracting this industry-wide information, TECHRETt-1 remains a strong predictor of focal firms returns next month. The magnitude and significance of the coefficient for TECHRETt-1 are virtually the same when we use industry-adjusted returns. Finally, consistent with the prior literature (Cohen and Lou, 2012), if the part of predictable returns of focal firms is solely attributable to delayed information processing, rather than industry-wide return continuation, we should see past industry returns have no predictive power for (RETt-INDRETt). Consistent with this prediction, we find that the coefficient on past industry returns, INDRETt-1, is indistinguishable from zero in column 4. The coefficients for the control variables are also consistent with prior literature: size, asset growth, and reversal variable are significantly negative related to future returns, while the coefficient of book-to-market, gross profitability, R&D intensity, and momentum are positive. 24

26 We further control for supplier and customer returns (Menzly and Ozbas, 2010), pseudo-conglomerate returns (Cohen and Lou, 2012), and stock turnover in Table 4. The results confirm the return predictability along the supply-chain as well as in the complicated firms, and low turnover stocks require higher future returns. We note that the magnitude and significance of the coefficient for TECHRETt-1 are qualitatively the same as our main results, indicating that the information diffused along the technological link cannot be explained by the information shocks from supply-chain, business segment, or the illiquidity of stocks. 5. Other Robustness Tests 5.1. Technology-linked return predictability across time In Online Appendix Table A.2, we examine whether the return predictability power of technology-linked firms varies across time. We divide our full sample periods into , , , and We then exactly repeat our baseline analysis from Table 3 for each subperiod. Our results hold up well to this time disaggregation. The coefficients of TECHRETt-1 are all positive and statistically significant after controlling for various return determinants. In fact, the only surprise in Online Appendix Table A.2 is that there appears to be little industry momentum in the last subperiod, which runs from The coefficient of INDRETt-1 is not significant for , while it is significant for the first three subperiods. It is hard to say whether this just reflects noise in a short sample or the fact that more arbitrageurs have caught on to the industry momentum 25

27 effects and are beginning to drive them out of existence. In any case, what is noteworthy from our perspective is that though the average degree of industry momentum may be declining over time, the technology momentum that we documented have been robust across four periods Persistence of technology closeness In this section, we examine the persistence, or stickiness, of technology closeness. More specifically, we examine the return predictability power when our technology momentum strategy is based on the lagged one-, two-, three-year technology closeness measures. Panel A of Online Appendix Table A.3 shows the correlations for TECHRETt-1, TECHRET_L1t-1, TECHRET_L2t-1, TECHRET_L3t-1. The results show that the correlation between TECHRETt-1 and its corresponding one-, two-, three-year lagged measures are remarkably positive and significant. For instance, the Pearson correlation coefficients between TECHRETt-1 and TECHRET_L1t-1 is When the lagged year increases, the correlation coefficient decreases, but the Pearson correlation between TECHRETt-1 and TECHRET_L3t-1 is still positive and significant, with the coefficient of In Panel B, we find that technology closeness is quite sticky, in that both TECHRET t-1 and its lagged forms predict focal firm returns, while the predictability power is lower, but still significant, for the lagged forms. Specifically, the one-year lagged TECHRET_L1t-1 generates equal weighted returns of 88 basis points per month (t=4.22), or roughly 10.6% per year. Controlling for other known return determinants generates equally good or even better results. Comparing to 26

28 TECHRETt-1, the return predictability power of TECHRET_L1t-1 is lower but still quite good. Results for TECHRET_L2t-1 and TECHRET_L3t-1 further confirm the return predictability of technology-linked returns, while predictability power decreases when the lagged year increases. But even the technology momentum strategy based on three-year lagged technology closeness measures works quite well, which means that firm locations in technology space are quite stable, and even for investors who do not have timely information on patent could still be able to make good return predictions based on this strategy Predictability for time-period beyond one month In Online Appendix Table A.4, we consider the profitability of (L, H) strategies following Moskowitz and Greenblatt (1999) to show the speed of information diffusion. In the (L, H) strategy, the technology momentum portfolios are formed based on L-month lagged returns, held for H months, and rebalanced monthly. Both equal-weighted and value-weighted results are reported for the (L, H) strategy of the hedge portfolio that, each month, buys (shorts) stocks with technology-linked returns in the highest (lowest) decile. For brevity, we only report the L = 1-, 3-, 6-, 12-month lagged and H = 1-, 6-, 12-, 24-, 36-month holding period strategies. Among the strategies that we consider, the short-term (1,1) strategy (i.e., L=1, H=1) is the most profitable. This result is robust to Daniel et al. (1997) (DGTW) characteristic-adjusted returns and industry-adjusted returns, which is consistent with prior regression results. Moreover, the profitability of short-term strategy is less sensitive to the length of ranking period L. For example, the equal-weighted raw 27