Heterogeneous Innovation over the Business Cycle*

Transcription

1 Heterogeneous Innovation over the Business Cycle* Gustavo Manso a, Benjamin Balsmeier b, and Lee Fleming a a) University of California, Berkeley, USA b) ETH Zurich, Switzerland Preliminary and Incomplete September 217 Abstract: Previous research has argued that innovative activities should be concentrated in recessions. However, innovation, as measured by R&D expenditure, seems to concentrate in booms. We argue that R&D expenditures do not capture the different dimensions of firms innovative search strategies. We introduce a model of innovative exploration and exploitation over the business cycle and present supporting evidence from a battery of patent-based measures. Exploitation strategies are more prevalent in booms while exploration strategies are more prevalent in recessions. Keywords: Exploration, Exploitation, Patents, Innovation, Business Cycles, Macroeconomic Shocks, Productivity, Growth * The authors thank Guan Cheng Li for invaluable research assistance. We gratefully acknowledge financial support from The Coleman Fung Institute for Engineering Leadership, the National Science Foundation (136228), and the Ewing Marion Kauffman Foundation. Errors and omissions remain the authors.

2 1. Introduction Macroeconomic shocks that induce business cycles may have an impact on innovative activities. Because innovation drives long-run growth, such shocks can have long-lived consequences, far beyond the duration of any particular cyclical episode. Understanding the effect of business cycles is thus crucial for macroeconomic policy and economic welfare. In recent years, economists have revived the notion, often associated with Joseph Schumpeter, that economic downturns play an important role in promoting long-term productivity. 1 According to this view, recessions are times of creative destruction, in which high innovation activity leads to enhancements in productivity and the demise of old technologies. The argument rests on the idea that the opportunity cost of innovative activities, i.e. the foregone sales that could be achieved instead, is lower in recessions. Another way to put it is that, during recessions, firms should focus on long-run investments since expected profits in the short run are low regardless. In contrast to this view, several studies have found that innovative activities, as measured by R&D expenditures and patents, have a procyclical bias. 2 Some explanations have been proposed, for example, based on credit constraints (Aghion et al. 27) or externalities in R&D (Barlevy 27). However, it remains a puzzle why innovative activities, at least using these measures, seem to be concentrated in booms and not in recessions. We argue that patent counts and R&D expenditures do not capture important dimensions of firms innovative search strategies. Following March (1991) and Manso (211), we model innovative search as a tension between exploration and exploitation. We measure the tension between exploration and exploitation with a battery of patent-based measures. To establish the validity of these measures, we show that they reliably predict capital and labor productivity as well as innovative success. Consistent with the model prediction, exploitation strategies are concentrated in booms while exploration strategies are concentrated in recessions. 1 See, for example, Cooper and Haltingwanger (1993), Caballero and Hammour (1994), Aghion and Saint-Paul (1998), and Canton and Uhlig (1999). 2 See, for example, Griliches (199), Geroski and Walters (1995), Fatas (2), Rafferty (23), Walde and Woitek (24), and Comin and Gertler (26). 2

3 These results suggest a more positive view on the welfare effects of macroeconomic fluctuations. If negative economic shocks encourage growth-enhancing exploration, economic recessions would tend to be shorter and less persistent than they would be otherwise. Cyclical fluctuations might even contribute positively to welfare if they allow the economy to balance exploration and exploitation. Innovation is the result of experimentation with new ideas. 3 The central tension that arises in experimentation is the tension between exploration and exploitation. Exploration involves search, risk-taking and experimentation with new technologies or new areas of knowledge. Exploitation, on the other hand, is the refinement of existing technologies. Because the opportunity cost of exploratory activities the additional output or sales that could have been achieved instead by a slightly refined product is lower in recessions, firms have incentives to undertake such activities in downturns. At the same time, during booms, firms have incentives to engage in exploitation, to avoid losing on the high sales of its current products. As a consequence, the model predicts that exploration is countercyclical while exploitation is procyclical. To measure exploration and exploitation we rely on patent data. Rather than just counting patents, we exploit the richness of the patent data to obtain a battery of measures that are intuitively related to exploration and exploitation. Our measures include the number of new technology classes entered by a firm, the number of patents in new technology classes, the number of patents in technology classes where the firm has previously patented, technological proximity, the number of backward citations, the number of self-citations, the number of claims, and the average age of inventors. We run a principal-component analysis with these measures. The first two components explain 79% of the joint variation of our eight patent measures. The single factor loadings support a theoretical model of exploration vs. exploitation, as the first component seems to load on measures associated with exploitation, while the second component seems to load on measures associated with exploration. We use the first component as our measure of exploitation and the second component as our measure of exploration. To establish the validity of these measures, we need to show that 3 See Arrow (1962). 3

4 exploration is indeed associated with breakthrough innovation and with future productivity and growth, while exploitation is associated with incremental innovation. We find that our measures of exploration and exploitation reliably predict a firm s future labor and capital productivity, sales in new-to-the-firm industries, as well as the type of innovation produced, as measured by whether its patents fall in the tails or in the middle of future citation distributions. We move on to study the innovative activity over the business cycle. Consistent with prior literature we find that R&D is strongly procyclical. Patent counts seem to be acyclical. Our main contribution is to break down innovation into exploration and exploitation. Using this more nuanced view of innovation, we find that exploration is countercyclical while exploitation is procyclical, as predicted by our simple model. Balsmeier, Fleming and Manso (217) use several of our patent-based innovation measures to study the effect of board independence on exploration and exploitation. Akcigit and Kerr (216) develop a growth model to analyze how different types of innovation contribute to economic growth and how the firm size distribution can have important consequences for the types of innovations realized. 2. Model We introduce a model of exploration and exploitation over the business cycle. The model is based on the simple two-armed bandit problem studied in Manso (211), but incorporates macroeconomic shocks. The economy exists for two periods. In each period, the representative firm in the economy takes either a well-known or a novel action. The well-known action has a known probability p of success. The novel action has an unknown probability q of success. The only way to learn about q is by taking the novel action. The expected probability q of success when taking the novel action is E[q] when the action is taken for the first time, E[q S] after experiencing a success with the novel action, and, E[q F] after experiencing a failure with the novel action. From Bayes rule, E[q F] < E[q] < E[q S]. 4

5 We assume that the novel action is of exploratory nature. This means that when the firm experiments with the novel action, it is initially not as likely to succeed as when it conforms to the conventional action. However, if the firm observes a success with the novel action, then the firm updates his beliefs about the probability q of success with the novel action, so that the novel action becomes perceived as better than the conventional action. This is captured as follows: E[q] < p < E[q S]. The macroeconomic state m can be either high (H) or low (L). If the macroeconomic state is currently m it remains in the same state next period with probability ρ m. Alternatively, it transitions into the other state n next period. Output in each period is given by ms in case of success and mf in case of failure. For simplicity, we assume risk-neutrality and zero discounting. There are only two action plans that need to be considered. The first relevant action plan, exploitation, is to take the well-known action in both periods. This action plan gives the following payoff if the macroeconomic state is m: pms + (1 p)mf + ρ m (pms + (1 p)mf) + (1 ρ m )(pns + (1 p)nf) The other relevant action plan, exploration, is to take the novel action in the first period and stick to it only if success is obtained. This action plan gives the following payoff it he macroeconomic state is m: E[q]mS + (1 E[q])mF + ρ m (E[q](E[q S]mS + (1 E[q S]))mF) + (1 E[q])(E[q F]mS + (1 E[q F])mF)) + (1 ρ m )(E[q](E[q S]nS + (1 E[q S])nF) + (1 E[q])(pnS + (1 p)nf)) The total payoff from exploration is higher than the total payoff from exploration iff: 5

6 E[q] m n(e[q S] p)(1 ρ m ) + m(1 + (E[q S] p)ρ m )) p If the firm tries the novel action, it obtains information about q. This information is useful for the firm s decision in the second period, since the firm can switch to the conventional action if it learns that the novel action is not worth pursuing. The coefficient multiplying p in the inequality above is less than 1. Therefore, the firm may thus be willing to try the novel action even though the initial expected probability E[q]of success with the novel action is lower than the probability p of success with the conventional work method. The coefficient multiplying p on the right-hand side of the inequality above is increasing in m and decreasing in n. Therefore the firm is more likely to explore in a recession (m = L, n = H) than in a boom (m = H, n = L). The intuition is that in a recession, the future is much more valuable than the present. Therefore, the firm is more forward-looking and is willing to explore for a larger set of parameters. If we add to the model some externality from exploring the novel action, there will be underexploration in equilibrium. Creating macroeconomic volatility may solve the underexploration problem as the greater the distance between H and L, the more likely a firm is to explore. Business cycle fluctuations may thus increase welfare in this model. 3. Data The empirical analysis is based on the joint availability of firm level accounting and market data of public US based firms in Compustat, disambiguated patent assignee data from Kogan et al. (KPSS, 217), the United States Patent and Trademark Office, and Fung Institute at UC Berkeley. We build firm level patent portfolios by aggregating eventually granted US patents from 1977 through 26 (results are robust over shorter periods) to the assignee-year level. 4 As we base our analysis on measures that have no obvious value in case of non-patenting activity, we only include firms in the analysis that applied for at least one patent in a given year. Table 1 shows the distribution of firm-year observations over the sampling period. 4 KPSS provide disambiguated assignee data for all patents granted by 21. However, the average patent remains with the USPTO about three years until it is granted. To avoid truncation influencing our results, we restrict the sample to years 27 and earlier. 6

7 Table 1 Frequency count of firm-year patent portfolio observations. Year Freq. Percent Cum , , , , , , , , , , , , , , , , , , , , Total 37,25 1 The number of observations increases over time as more and more firms begin to patent over the sampling period. In order to limit selection in and out of the sample only firms that appear at least 4 times in the data are included in the analysis (results remain robust to 3 or 5 years as the threshold value). Next to the more common variables, e.g. total number of patents applied for in a given year and total number of future citations that those patents received, we report additional measures that enable us to assess the direction of innovation pursued by companies. The analysis below demonstrates that a differentiation between innovative search in terms of exploration and exploitation has crucial implications for future innovative success, as well as other important outcomes, e.g. capital and labor productivity. Many of the measures rely on the USPTO s 7

8 classification of all patents into one of approximately 13 CPC technology classes. The measures are firm-year level aggregations based on information from each patent document, where year is the application year of each patent. This comes closer to the time the invention was actually made. The analyses use the following patent portfolio characteristics: 1. Number of patents that are filed in a 3-digit technology classes where the given firm has filed no other patent beforehand in that class. 2. Number of patents that are filed in a 3-digit technology classes where the given firm has filed another patent beforehand in that class. 3. Number of new technology classes entered where the given firm has filed no patent beforehand in that class. 4. Technological proximity between the patents filed in year t and the existing patent portfolio held by the same firm up to year t-1 (the normalized correlation between two years of the proportion of activity in a given class, calculated according to Jaffe 1989). 5. Number of prior art citations to other patents ( backward citations ). 6. Number of prior art citations to patents held by the same firm ( self-backward citations ). 7. Number of claims a patent makes. 8. Average age of the inventor(s) mentioned on the patent document as calculated by the time difference between the first time an inventor occurs in the Fung Institute s patent database and the application year of a given patent. In order to assess the innovative success of patent portfolios we break out the number of patents filed in a given year according to their place in the distribution of future citations by all patents granted until October 216 within a given technology class and application year. Here we report 6 measures: 1. Number of patents that fall into the top 1% of the future citation distribution within a given technology class and application year. 2. Number of patents that fall into the top 5 % to top 1% of the future citation distribution within a given technology class and application year. 3. Number of patents that fall into the top 1 to top 5% of the future citation distribution within a given technology class and application year. 8

9 4. Number of patents that fall into the top 25 to top 1% of the future citation distribution within a given technology class and application year. 5. Number of patents that fall into the top 5 to top 25% of the future citation distribution within a given technology class and application year. 6. Number of patents that received no citation by any granted patent until October 216. Table 2 reports summary statistics on all these variables, and Table 3 shows the correlation matrix. 9

10 Table 2 Summary statistics Variable N mean Median sd min max Patents Top 1% Top 5-2% Top 1-6% Top 25-11% Top 5-26% No cite All future cites New tech classes entered Patents in new classes Patents in known classes Technological proximity Av. age of inventors Backward citations Self-citations Claims Patent stock Notes: This table reports summary statistics of variables used in the study. Patents is the total number of eventually granted patents applied for in a given year. Top 1% are the number of patents that fall into the 1% most cited patents within a given 3- digit technology class and application year. Top 5% to 2% are the number of patents that fall into the 5% to 2% most cited patents within a given 3-digit technology class and application year. Top 1% to 6% are the number of patents that fall into the 1% to 6% most cited patents within a given 3-digit technology class and application year. Top 25% to 11% are the number of patents that fall into the 25% to 11% most cited patents within a given 3-digit technology class and application year. Top 5% to 26% are the number of patents that fall into the 5% to 26% most cited patents within a given 3-digit technology class and application year. No cite are the number of patents that are not cited by any other patent. All future cites is the total number of future citations. New classes entered is the number of technology classes where a firm filed at least one patent but no other patent beforehand. Patents in new/known classes is the number of patents that are filed in classes where the given firm has filed no/at least one other patent beforehand. Technological proximity is the technological proximity between the patents filed in year t to the existing patent portfolio held by the same firm up to year t-1, calculated according to Jaffe (1989). Av. Age of inventors the average time difference between the first time an inventor occurs in the Fung Institute s patent database and the application year of a given patent. Backward citations is the total number of citations made to other patents. Self-citations is the total number of cites to patents held by the same firm. Claims is the total number of claims on each patent. Patent stock is the sum of all patents held by a given firm up to the year t-1. 1

11 Table 3 - Correlation matrix (1) Patents 1. (2) Top 1% (3) Top 5-2% (4) Top 1-6% (5) Top 25-11% (1) (2) (3) (4) (5) (6) (7) (8) (9) (1) (11) (12) (13) (14) (15) (16) (6) Top 5-26% (7) No cite (8) All future cites (9) New tech classes entered (1) Patents in new classes (11) Patents in known classes (12) Technological proximity (13) Av. age of inventors (14) Backward citations (15) Self-citations (16) Claims (17) Patent stock Notes: This table reports pairwise correlations of the log-transformed variables used in the study. Patents is the total number of eventually granted patents applied for in a given year. Top 1% are the number of patents that fall into the 1% most cited patents within a given 3-digit technology class and application year. Top 5% to 2% are the number of patents that fall into the 5% to 2% most cited patents within a given 3-digit technology class and application year. Top 1% to 6% are the number of patents that fall into the 1% to 6% most cited patents within a given 3-digit technology class and application year. Top 25% to 11% are the number of patents that fall into the 25% to 11% most cited patents within a given 3-digit technology class and application year. Top 5% to 26% are the number of patents that fall into the 5% to 26% most cited patents within a given 3-digit technology class and application year. No cite are the number of patents that are not cited by any other patent. All future cites is the total number of future citations. New classes entered is the number of technology classes where a firm filed at least one patent but no other patent beforehand. Patents in new/known classes is the number of patents that are filed in classes where the given firm has filed no/at least one other patent beforehand. Technological proximity is the technological proximity between the patents filed in year t to the existing patent portfolio held by the same firm up to year t-1, calculated according to Jaffe (1989). Av. Age of inventors the average time difference between the first time an inventor occurs in the Fung Institute s patent database and the application year of a given patent. Backward citations is the total number of citations made to other patents. Self-citations is the total number of cites to patents held by the same firm. Claims is the total number of claims on each patent. Patent stock is the sum of all patents held by a given firm up to the year t PCA Analysis In order to classify the patent portfolios according to their exploration and exploitation focus we run a PCA analysis (similar results are obtained with a count based approach, or running a similar PCA at the patent level). The commonly taken threshold of Eigenvalues above one is only passed by two components, suggesting that extracting two components is the best way to proceed. It supports the theoretical focus on two dimensions. The output shown in Tables 4 to 5 indicate that 78 percent of the joint variation of our eight patent variables of interest can be explained by the first two principal components. In order to 11

12 optimize the factor loadings we apply a Varimax rotation of the two extracted components (results are robust to other orthogonal rotations and not applying any rotation). Table 4 shows the corresponding results. Table 5 shows how much and in which direction each variable loads on the two components. Loadings below.2 are not shown for easier comparability. The single loadings support a theoretical model of exploration vs. exploitation. Patents in known classes, technological proximity, inventor age, backward citations, self-backward citations, and claims all positively load on component 1. Hence, we label component 1 exploitation. The number of new technology classes entered and patents in new to the firm technology classes strongly and positively load on component two. Negatively related to component two is the technological proximity and the age of the inventors. We label this component exploration. Measures of originality and generality (Hall, Jeffe and Trajtenberg 21 on whether the patent cites a wide variety of classes or is cited in turn by a wide variety) do not load on either of our components (neither at the firm nor patent level). The measures do not map clearly to our theory; a patent could cite a wide variety of classes that had never been cited together before, or had been heavily cited together before. In other words, a highly original patent could be citing a previously uncombined set of classes or a very commonly combined set of classes. Adding the measures to our citation regressions below do not alter the results, and the coefficient of originality remains close to zero in terms of size and statistical significance. 12

13 Tables 4 and 5 Principal Component Analysis Component Variance Difference Proportion Cumulative Comp Comp Notes: This table reports the results of a Principal Component Analysis after Varimax Rotation. Only components with Eigenvalues above one are extracted. The 8 variables that entered the PCA are: new classes entered, patents in new/known classes, technological proximity, av. age of inventors, backward citations, self-citations, and claims; all variables log-transformed. Variable Comp1 Comp2 Unexplained New tech classes entered.63.7 Patents in new classes.63.7 Patents in known classes.48.8 Technological proximity Backward citations Self-citations Claims.46.9 Av. age of inventors Notes: This table reports the results of a Principal Component Analysis after Varimax Rotation. Only components with Eigenvalues above one are extracted. All variables log-transformed. Variable definitions provided above. Table 6 KMO test Variable KMO New tech classes entered.57 Patents in new classes.56 Patents in known classes.76 Technological proximity.74 Backward citations.9 Self-citations.86 Claims.83 Av. age of inventors.8 Overall.76 Notes: This table reports the Kaiser-Mayer- Olkin (KMO) measures on sampling adaquacy. All variables log-transformed. Variable definitions provided above. 13

14 The Kaiser-Mayer-Olkin measure of sampling adequacy, shown in Table 6, confirms that the data can be summarized using a PCA analysis. The correlation between the two factors is.37. While this correlation indicates that there are some firms working in areas that score high on exploration and exploitation, the correlation is far from being perfect, implying substantial independent variation. For instance, some firms enter new tech classes with relatively young inventors while at the time also filing new patents in known tech areas with older inventors. This is illustrated by Figure 1 that plots the factor values of the exploration component against the factor values of the exploitation component. The lines show the median values of each component that divides the data into four categories. Exploration scores around zero occur rarely because of the discontinuity that comes from patenting in new to the firm technology classes or not and because of the relatively broad CPC technology classes that the USPTO started to use in 216 and will continue to use in the future. As a robustness check we re-ran all analysis based on the finer but recently abandoned technology classes originally assigned to patents by the USPTO and find qualitatively the same results. Figure 1 Scatter Plot of PCA scores Notes: This graph plots the component scores of Exploration and Exploitation extracted from the Principal Component Analysis shown above. Red lines mark the median values of each factor. 19% of the observations are each in the upper left and lower right quadrants, 31% in each of the other quadrants. 14

15 Table 7 shows regressions of: (1) Number of patents in top 1% future citation distribution, (2) Number of patents in top 5%, but not including top 1%, (3) Number of patents in top 1% distribution, but not including top 5%, (4) Number of patents in top 25% distribution, but not including top 1%, (5) Number of patents in top 5% distribution, but not including top 25%, (6) Number of patents that received no citation at all, (7) Number of all future citations. Three main explanatory variables correspond to the mutually exclusive quadrants in Figure 1: Exploit, Explore and Exploit + Explore. These indicators categorize the search strategy in a particular year as observed by a given firm s patent portfolio. We further control for the firms patent stock, R&D investment per total assets, firm age, firm size and time and industry fixed effects (the latter not shown). The regressions illustrate significant explanatory power, based on a firm s patent portfolio score and exploration vs. exploitation strategy. Firms that score low on both exploration and exploitation patent very little and provide a low technology category baseline comparison. Consistent with predictions from March (1991), firms with exploitation portfolios garner future citations from the middle of the distribution; they receive significantly fewer citations from the highly cited and not at all cited tails and significantly more from the middle of the distribution. Firms with exploration portfolios receive significantly more citations except from the bottom of the distribution, though the effect sizes are not big. Exploitation firms receive more total citations as well. The firms with a balanced strategy garner significantly more citations than focused strategies across the distribution; the magnitude peaks for patents right above the mean number of citations. Balanced strategy firms also receive more citations on average, while exploring firms find themselves more likely with patents that are not cited at all or highly cited patents, resulting in a lower number of cites on average. This is also illustrated by Figure 2 that plots the densities of the average number of cites to patents field by firms that focus on exploitation as compared to patents field by firms focusing on exploration. 15

16 Dependent variable Top 1% Top 5-2% Table 7 - Patent citation distribution break out a b c d e f g Top 1-6% Top 25-11% Top 5-26% Not cited Av. future cites log(patent stock).127***.243***.283***.367***.426***.31*** -.13* (.9) (.12) (.12) (.13) (.12) (.12) (.8) R&D.164***.3***.34***.389***.458***.24***.28*** (.27) (.39) (.41) (.46) (.41) (.34) (.65) log(age) -.95*** -.18*** -.217*** -.282*** -.311*** -.23*** -.18*** (.8) (.12) (.12) (.13) (.13) (.13) (.13) log(total assets).21***.52***.61***.89***.124***.79*** -.14** (.4) (.5) (.6) (.6) (.6) (.5) (.7) Exploit -.65***.2.49**.321***.55*** -.145***.355*** (.14) (.19) (.19) (.2) (.19) (.19) (.19) Explore.81***.16***.195***.258***.32***.21*** -.27 (.7) (.1) (.1) (.12) (.12) (.1) (.17) Exploit + Explore.17***.38***.431***.825*** 1.82***.193***.346*** (.12) (.16) (.17) (.19) (.19) (.17) (.19) Time + Industry FE Yes Yes Yes Yes Yes Yes Yes N R Notes: This tables presents OLS regression results. All dependent variables are log-transformed. Top 1% are the number of patents that fall into the 1% most cited patents within a given 3-digit technology class and application year. Top 5% to 2% are the number of patents that fall into the 5% to 2% most cited patents within a given 3-digit technology class and application year. Top 1% to 6% are the number of patents that fall into the 1% to 6% most cited patents within a given 3-digit technology class and application year. Top 25% to 11% are the number of patents that fall into the 25% to 11% most cited patents within a given 3-digit technology class and application year. Top 5% to 26% are the number of patents that fall into the 5% to 26% most cited patents within a given 3-digit technology class and application year. No cite are the number of patents that are not cited by any other patent. All future cites is the total number of future citations. R&D is R&D expenditures scaled by total assets. Age is years since IPO. Exploit / Explore indicates all firms focusing on exploitation/exploration, as classified by the PCA shown above. Exploit+Explore indicates all firms that score high (>median) on both components. All categories are mutually exclusive. Time fixed effects are 23 dummies for each year. Industry fixed effects are 25 dummies for each 2-digit- SIC industry. All models include an intercept which is omitted in the table. Standard errors clustered at the firm level appear in parentheses. ***, ** and * indicate a significance level of 1%, 5%, and 1%, respectively. 16

17 Figure 2 Average future cites distribution Average cites per patent Density log(future cites per patent) Exploiting Firms kernel = epanechnikov, bandwidth =.3 Exploring Firms Notes: This graph shows density plots of the logarithm of the average amount of future cites per patent for firms that focus Exploration and Exploitation as defined by the PCA analysis above. Only firms with min 1 patents in a given year. Performance regressions Next we run regressions of (1) capital productivity, (2) labor productivity, (3) sales in new industries, and (4) the similarity of firm s product portfolios with other firms product portfolios (Hoberg and Phillips, 215, 21), on the patent portfolio classification indicators and the previously used controls. Capital productivity is defined as the log ratio of firm s total sales plus change in inventories divided by capital. Labor productivity is the same but scaled by number of employees instead of capital. 5 In regressions of productivity of capital on our measures of exploration and exploitation, the number of employees serves as a control for firm size, while in regressions of labor productivity total assets serve a proxy for firm size to avoid spurious correlations due to equal denominators of the dependent variable. All explanatory variables are lagged by one year to allow for a time lag between actual innovation and the realization of its effect on output. Firms pursuing exploitation strategies experience a positive correlation with capital productivity, an insignificant correlation with labor productivity, and a negative correlation with sales in new industries. Firms pursuing exploration strategies experience positive correlations with both productivities and with sales in new industries. Balanced firms 5 Variable definitions follow Kaplan et al. (215). 17

18 experience positive and significant correlations with both productivities but no significant correlation with sales in new industries. The market entry results are resembled, when we measure the competition in product markets based on Hoberg and Phillips textual analysis of firms 1k filings. Note that all regressions include controls for patent stock: they do not simply reflect high R&D investment and high technology firms. Table 8 Productivity, Market Entry and Product Proximity Regressions Dependent variable a b c d Labor Sales in Productivity new Ind. Capital Productivity Prod. Proximity log(patent stock) t-1.25***.3***.72*** -.29*** (.9) (.9) (.22) (.6) R&D t *** -.977***.31***.713*** (.74) (.95) (.87) (.52) log(age) t ***.12*** *** (.14) (.15) (.31) (.1) log(total assets) t-1.9***.236***.43*** (.8) (.15) (.5) log(employees) t-1.984*** (.7) Exploit t-1.43** ***.52*** (.18) (.21) (.43) (.15) Explore t-1.74***.125***.122*** -.3** (.14) (.17) (.29) (.12) Exploit + Explore t-1.13***.127*** (.18) (.2) (.45) (.14) Time + Industry FE Yes Yes Yes Yes N R Notes: This tables presents OLS regression results. All dependent variables are logtransformed and winsorized by year at the 1% level. Sales in new industries is the log of the average sales generated in a 3-digit SIC industry in t+2 to t+4, where the given firm has not generated any sales beforehand. Model (d) is a regression of product proximity based on textual analysis of firms 1k fillings by Hoberg and Phillips (215, 21), multiplied by 1 to give it a proportional value. The number of observations is lower in (d) because the proximity measure is only available since Patent stock are all eventually granted patents applied for up to year t-1. R&D is R&D expenditures scaled by total assets. Age is years since IPO. Exploit / Explore indicates all firms focusing on exploitation/ exploration, as classified by the PCA shown above. Exploit+Explore indicates all firms that score high (>median) on both components. All categories are mutually exclusive. Time fixed effects are 29 dummies for each year. Industry fixed effects are 25 dummies for each 2-digit-SIC industry. All models include an intercept which is omitted in the table. Standard errors clustered at the firm level appear in parentheses. ***, ** and * indicate a significance level of 1%, 5%, and 1%, respectively. Innovation over the business cycle 18

19 The next graphs illustrate the evolution of firms R&D expenditure and patent counts over the business cycle, i.e. the time between the first year after a trough and the following peak. Business cycle data and definitions of trough and peak years are taken from the NBER. 6 Regarding Figure 2 below, coefficients for each year are derived from regressing the log of a firm s R&D expenditures on dummies for each business cycle year between a trough and the following peak, 25 SIC 2-digit industry dummies, and three separate dummies for each cycle. Contraction years serve as the reference category. Coefficient sizes should be interpreted as relative changes compared to the average cycle level. Figure 3 illustrates results from the same model but with the log of the number of patents as the dependent variable. 6 See 19

20 Figure 3 R&D expenditures R&D Investment coefficient size Years between trough and peak b-coefficients 95%-confidence-interval Figure 4 Number of patents Number of Patents coefficient size Years between trough and peak b-coefficients 95%-confidence-interval Another graph below is derived from the same regressions but with our exploration and exploitation factors as the dependent variables instead of the log of the R&D expenditures and the log of total number of patents. As predicted by the model, the graphs show that exploration is countercyclical while exploitation is procyclical. Figure 6 below shows the estimated values separately for each cycle covered by the data. It shows that the predicted cyclical change in exploration and exploitation is present in all cycles. Further, Table 9 shows that all findings are supported by regressions of yearly changes in the exploitation (and exploration) scores per firm on yearly GDP growth and yearly sales growth 2

21 per industry. Results are even more pronounced, when only frequently patenting firms are considered ( 15 years in the dataset, see Figure A1 in the Appendix). Figure 5 Exploration and Exploitation over the Business Cycle Exploration and Exploitation Change in Exploitation/Exploration Years between trough and peak Exploitation Exploration 95%-confidence-band Dependent variable Table 9 Exploration, Exploitation, and Growth Δ Exploitation score Δ Exploration score Δ Exploitation score Δ Exploration score log(patent stock) t ***.74*** -.116***.74*** (.5) (.1) (.5) (.1) R&D t-1.118*** -.261***.19*** -.243*** (.3) (.76) (.31) (.76) log(age) t ***.97*** -.47***.95*** (.7) (.8) (.7) (.8) log(total assets) t-1.58*** -.38***.57*** -.37*** (.5) (.6) (.5) (.6) Sales Growth.15* -.137* (.6) (.7) GDP Growth *** (.358) (.255) Industry FE Yes Yes Yes Yes N R Notes: This table presents OLS regression of changes in firms exploration and exploitation scores from t-1 to t. Exploitation and exploration scores are derived from the PCA shown above. GDP growth is the fractional change of GDP from t-1 to t. Sales growth is the fractional change of sales in a given firm s 3-digit industry classification from t-1 to t. Patent stock are all eventually granted patents applied for up to year t-1. R&D is R&D expenditures scaled by total assets. Age is years since IPO. Industry fixed effects are 25 dummies for each 2-digit-SIC industry. All models include an intercept which is omitted in the table. Standard errors clustered at the firm level appear in parentheses. ***, ** and * indicate a significance level of 1%, 5%, and 1%, respectively. 21

22 5. Conclusion Previous research argues that innovative activities should be countercyclical. However, it finds that common measures of innovation, R&D expenditure and patent counts, are procyclical. We provide a solution to this puzzle by modelling innovative search as a tension between exploration and exploitation. We rely on a battery of patent-based measures to separate exploration and exploitation. Consistent with the model, exploitation strategies are procyclical while exploration strategies are countercyclical. References Aghion, Philippe and Gilles Saint Paul, Virtues of Bad Times: Interaction between Productivity Growth and Economic Fluctuations Macroeconomic Dynamics, September, 2(3), p Aghion, Philippe, Philippe Askenazy, Nicolas Berman, Gilbert Cette, 212. Credit Constraints and the Cyclicality of R&D Investment: Evidence from France Journal of the European Economic Association 1(5), p Akcigit, Ufuk and William Kerr, 216. Growth Through Heterogenenous Innovations, Journal of Political Economy (forthcoming). Arrow, Kenneth, Economic Welfare and the Allocation of Resources for Invention in The Rate and Direction of Inventive Activity: Economic and Social Factors, edited by Richard Nelson, Princeton, NJ: Princeton University Press, p Balsmeier, Benjamin, Lee Fleming, and Gustavo Manso, 217. Independent Boards and Innovation, Journal of Financial Economics vol. 123, Barlevy, Gadi, 27. On the Cyclicality of Research and Development American Economic Review, 97(4), p

23 Canton, Eric and Harald Uhlig, Growth and the Cycle: Creative Destruction versus Entrenchment Journal of Economics, 69(3), p Comin, Diego and Mark Gertler, 26. Medium-Term Business Cycles American Economic Review, September, 96(3), June, p Cooper, Russell and John Haltiwanger, The Aggregate Implications of Machine Replacement: Theory and Evidence American Economic Review, June, 83(3), p Fatas, Antonio. 2. Do Business Cycles Cast Long Shadows? Short-Run Persistence and Economic Growth. Journal of Economic Growth, 5(2): Geroski, Paul A., and C. F. Walters Innovative Activity over the Business Cycle. Economic Journal, 15(431): Griliches, Zvi Patent Statistics as Economic Indicators: A Survey. Journal of Economic Literature, 28(4): Kogan, L., Papanikolaou, A., Seru, A., Stoffman, N Technological Innovation, Resource Allocation and Growth. Forthcoming: Quarterly Journal of Economics. Manso, Gustavo Motivating Innovation. Journal of Finance, 66(5), p March, James Exploration and Exploitation in Organizational Learning Organization Science, 2(1), p Rafferty, Matthew C. 23. Do Business Cycles Influence Long-Run Growth? The Effect of Aggregate Demand on Firm-Financed R&D Expenditures. Eastern Economic Journal, 29(4): Walde, Klaus, and Ulrich Woitek. 24. R&D Expenditure in G7 Countries and the Implications for Endogenous Fluctuations and Growth. Economics Letters, 82(1):

24 24

25 Appendix Figure A1: Only frequently patenting firms (observed 15 years) Exploration and Exploitation Change in Exploitation/Exploration Years between trough and peak Exploitation Exploration 95%-confidence-band Exploration and Exploitation per Business Cycle Explore and Exploit Change in Exploitation/Exploration Years between trough and peak Exploitation Exploration

26 Single measures (full sample) Claims Backward Self-Cites coefficient size coefficient size Years between trough and peak Years between trough and peak b-coefficients 95%-confidence-interval b-coefficients 95%-confidence-interval coefficient size Backward Cites coefficient size Tech Proximity Years between trough and peak Years between trough and peak b-coefficients 95%-confidence-interval b-coefficients 95%-confidence-interval Patents in known tech class Patents in new tech class coefficient size coefficient size Years between trough and peak Years between trough and peak b-coefficients 95%-confidence-interval b-coefficients 95%-confidence-interval Inventor Tenure Sales in new industries coefficient size coefficient size Years between trough and peak Years between trough and peak b-coefficients 95%-confidence-interval b-coefficients 95%-confidence-interval 26